Merge pull request #1 from PaddlePaddle/dygraph

Dygraph

.gitignore

@@ -24,4 +24,8 @@ output/
 
 build/
 dist/
-paddleocr.egg-info/
\ No newline at end of file
+paddleocr.egg-info/
+/deploy/android_demo/app/OpenCV/
+/deploy/android_demo/app/PaddleLite/
+/deploy/android_demo/app/.cxx/
+/deploy/android_demo/app/cache/

MANIFEST.in

@@ -4,4 +4,5 @@ include README.md
 recursive-include ppocr/utils *.txt utility.py logging.py
 recursive-include ppocr/data/ *.py
 recursive-include ppocr/postprocess *.py
-recursive-include tools/infer *.py
\ No newline at end of file
+recursive-include tools/infer *.py
+recursive-include ppocr/utils/e2e_utils/ *.py
\ No newline at end of file

PPOCRLabel/PPOCRLabel.py

@@ -147,6 +147,7 @@ class MainWindow(QMainWindow, WindowMixin):
         self.itemsToShapesbox = {}
         self.shapesToItemsbox = {}
         self.prevLabelText = getStr('tempLabel')
+        self.noLabelText = getStr('nullLabel')
         self.model = 'paddle'
         self.PPreader = None
         self.autoSaveNum = 5
@@ -1020,7 +1021,7 @@ class MainWindow(QMainWindow, WindowMixin):
         item.setText(str([(int(p.x()), int(p.y())) for p in shape.points]))
         self.updateComboBox()
 
-    def updateComboBox(self): # TODO：貌似没用
+    def updateComboBox(self):
         # Get the unique labels and add them to the Combobox.
         itemsTextList = [str(self.labelList.item(i).text()) for i in range(self.labelList.count())]
 
@@ -1040,7 +1041,7 @@ class MainWindow(QMainWindow, WindowMixin):
             return dict(label=s.label,  # str
                         line_color=s.line_color.getRgb(),
                         fill_color=s.fill_color.getRgb(),
-                        points=[(p.x(), p.y()) for p in s.points],  # QPonitF
+                        points=[(int(p.x()), int(p.y())) for p in s.points],  # QPonitF
                        # add chris
                         difficult=s.difficult)  # bool
 
@@ -1069,7 +1070,7 @@ class MainWindow(QMainWindow, WindowMixin):
             # print('Image:{0} -> Annotation:{1}'.format(self.filePath, annotationFilePath))
             return True
         except:
-            self.errorMessage(u'Error saving label data')
+            self.errorMessage(u'Error saving label data', u'Error saving label data')
             return False
 
     def copySelectedShape(self):
@@ -1802,10 +1803,14 @@ class MainWindow(QMainWindow, WindowMixin):
                     result.insert(0, box)
                     print('result in reRec is ', result)
                     self.result_dic.append(result)
-                    if result[1][0] == shape.label:
-                        print('label no change')
-                    else:
-                        rec_flag += 1
+                else:
+                    print('Can not recognise the box')
+                    self.result_dic.append([box,(self.noLabelText,0)])
+
+                if self.noLabelText == shape.label or result[1][0] == shape.label:
+                    print('label no change')
+                else:
+                    rec_flag += 1
 
             if len(self.result_dic) > 0 and rec_flag > 0:
                 self.saveFile(mode='Auto')
@@ -1836,9 +1841,14 @@ class MainWindow(QMainWindow, WindowMixin):
                     print('label no change')
                 else:
                     shape.label = result[1][0]
-                self.singleLabel(shape)
-                self.setDirty()
-            print(box)
+            else:
+                print('Can not recognise the box')
+                if self.noLabelText == shape.label:
+                    print('label no change')
+                else:
+                    shape.label = self.noLabelText
+            self.singleLabel(shape)
+            self.setDirty()
 
     def autolcm(self):
         vbox = QVBoxLayout()

PPOCRLabel/README.md

@@ -29,9 +29,7 @@ PaddleOCR models has been built in PPOCRLabel, please refer to [PaddleOCR instal
 
 ### 2. Install PPOCRLabel
 
-#### Windows + Anaconda
-
-Download and install [Anaconda](https://www.anaconda.com/download/#download) (Python 3+)
+#### Windows
 
 ```
 pip install pyqt5

PPOCRLabel/README_ch.md

@@ -31,7 +31,7 @@ PPOCRLabel是一款适用于OCR领域的半自动化图形标注工具，内置P
 PPOCRLabel内置PaddleOCR模型，故请参考[PaddleOCR安装文档](https://github.com/PaddlePaddle/PaddleOCR/blob/develop/doc/doc_ch/installation.md)准备好PaddleOCR，并确保PaddleOCR安装成功。
 
 ### 2. 安装PPOCRLabel
-#### Windows + Anaconda
+#### Windows
 
 ```
 pip install pyqt5

PPOCRLabel/libs/canvas.py

@@ -45,7 +45,7 @@ class Canvas(QWidget):
     CREATE, EDIT = list(range(2))
     _fill_drawing = False # draw shadows
 
-    epsilon = 11.0
+    epsilon = 5.0
 
     def __init__(self, *args, **kwargs):
         super(Canvas, self).__init__(*args, **kwargs)

PPOCRLabel/libs/utils.py

@@ -124,6 +124,15 @@ def natural_sort(list, key=lambda s:s):
 
 
 def get_rotate_crop_image(img, points):
+    # Use Green's theory to judge clockwise or counterclockwise
+    # author: biyanhua
+    d = 0.0
+    for index in range(-1, 3):
+        d += -0.5 * (points[index + 1][1] + points[index][1]) * (
+                    points[index + 1][0] - points[index][0])
+    if d < 0: # counterclockwise
+        tmp = np.array(points)
+        points[1], points[3] = tmp[3], tmp[1]
 
     try:
         img_crop_width = int(

PPOCRLabel/resources/strings/strings-zh-CN.properties

@@ -87,6 +87,7 @@ creatPolygon=四点标注
 drawSquares=正方形标注
 saveRec=保存识别结果
 tempLabel=待识别
+nullLabel=无法识别
 steps=操作步骤
 choseModelLg=选择模型语言
 cancel=取消

PPOCRLabel/resources/strings/strings.properties

@@ -77,7 +77,7 @@ IR=Image Resize
 autoRecognition=Auto Recognition
 reRecognition=Re-recognition
 mfile=File
-medit=Eidt
+medit=Edit
 mview=View
 mhelp=Help
 iconList=Icon List
@@ -87,6 +87,7 @@ creatPolygon=Create Quadrilateral
 drawSquares=Draw Squares
 saveRec=Save Recognition Result
 tempLabel=TEMPORARY
+nullLabel=NULL
 steps=Steps
 choseModelLg=Choose Model Language
 cancel=Cancel

README.md

@@ -32,7 +32,8 @@ PaddleOCR supports both dynamic graph and static graph programming paradigm
 
 <div align="center">
     <img src="doc/imgs_results/ch_ppocr_mobile_v2.0/test_add_91.jpg" width="800">
-    <img src="doc/imgs_results/ch_ppocr_mobile_v2.0/00018069.jpg" width="800">
+    <img src="doc/imgs_results/multi_lang/img_01.jpg" width="800">
+    <img src="doc/imgs_results/multi_lang/img_02.jpg" width="800">
 </div>
 
 The above pictures are the visualizations of the general ppocr_server model. For more effect pictures, please see [More visualizations](./doc/doc_en/visualization_en.md).
@@ -42,7 +43,7 @@ The above pictures are the visualizations of the general ppocr_server model. For
 - Scan the QR code below with your Wechat, you can access to official technical exchange group. Look forward to your participation.
 
 <div align="center">
-<img src="https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/release/2.0/doc/joinus.PNG"  width = "200" height = "200" />
+<img src="https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/dygraph/doc/joinus.PNG"  width = "200" height = "200" />
 </div>
 
 

README_ch.md

@@ -8,9 +8,9 @@ PaddleOCR同时支持动态图与静态图两种编程范式
 - 静态图版本：develop分支
 
 **近期更新**
+- 2021.4.8 release 2.1版本，新增AAAI 2021论文[端到端识别算法PGNet](./doc/doc_ch/pgnet.md)开源，[多语言模型](./doc/doc_ch/multi_languages.md)支持种类增加到80+。
 - 2021.2.1 [FAQ](./doc/doc_ch/FAQ.md)新增5个高频问题，总数162个，每周一都会更新，欢迎大家持续关注。
-- 2021.1.26,28,29 PaddleOCR官方研发团队带来技术深入解读三日直播课，1月26日、28日、29日晚上19:30，[直播地址](https://live.bilibili.com/21689802)
-- 2021.1.21 更新多语言识别模型，目前支持语种超过27种，[多语言模型下载](./doc/doc_ch/models_list.md)，包括中文简体、中文繁体、英文、法文、德文、韩文、日文、意大利文、西班牙文、葡萄牙文、俄罗斯文、阿拉伯文等，后续计划可以参考[多语言研发计划](https://github.com/PaddlePaddle/PaddleOCR/issues/1048)
+- 2021.1.21 更新多语言识别模型，目前支持语种超过27种，包括中文简体、中文繁体、英文、法文、德文、韩文、日文、意大利文、西班牙文、葡萄牙文、俄罗斯文、阿拉伯文等，后续计划可以参考[多语言研发计划](https://github.com/PaddlePaddle/PaddleOCR/issues/1048)
 - 2020.12.15 更新数据合成工具[Style-Text](./StyleText/README_ch.md)，可以批量合成大量与目标场景类似的图像，在多个场景验证，效果明显提升。
 - 2020.11.25 更新半自动标注工具[PPOCRLabel](./PPOCRLabel/README_ch.md)，辅助开发者高效完成标注任务，输出格式与PP-OCR训练任务完美衔接。
 - 2020.9.22 更新PP-OCR技术文章，https://arxiv.org/abs/2009.09941
@@ -46,7 +46,7 @@ PaddleOCR同时支持动态图与静态图两种编程范式
 - 微信扫描二维码加入官方交流群，获得更高效的问题答疑，与各行各业开发者充分交流，期待您的加入。
 
 <div align="center">
-<img src="https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/release/2.0/doc/joinus.PNG"  width = "200" height = "200" />
+<img src="https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/dygraph/doc/joinus.PNG"  width = "200" height = "200" />
 </div>
 
 ## 快速体验
@@ -74,11 +74,13 @@ PaddleOCR同时支持动态图与静态图两种编程范式
 ## 文档教程
 - [快速安装](./doc/doc_ch/installation.md)
 - [中文OCR模型快速使用](./doc/doc_ch/quickstart.md)
+- [多语言OCR模型快速使用](./doc/doc_ch/multi_languages.md)
 - [代码组织结构](./doc/doc_ch/tree.md)
 - 算法介绍
     - [文本检测](./doc/doc_ch/algorithm_overview.md)
     - [文本识别](./doc/doc_ch/algorithm_overview.md)
-    - [PP-OCR Pipline](#PP-OCR)
+    - [PP-OCR Pipeline](#PP-OCR)
+    - [端到端PGNet算法](./doc/doc_ch/pgnet.md)
 - 模型训练/评估
     - [文本检测](./doc/doc_ch/detection.md)
     - [文本识别](./doc/doc_ch/recognition.md)
@@ -112,7 +114,7 @@ PaddleOCR同时支持动态图与静态图两种编程范式
 
 
 <a name="PP-OCR"></a>
-## PP-OCR Pipline
+## PP-OCR Pipeline
 <div align="center">
     <img src="./doc/ppocr_framework.png" width="800">
 </div>

configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml

@@ -7,11 +7,6 @@ Global:
   save_epoch_step: 1200
   # evaluation is run every 5000 iterations after the 4000th iteration
   eval_batch_step: [3000, 2000]
-  # 1. If pretrained_model is saved in static mode, such as classification pretrained model
-  #    from static branch, load_static_weights must be set as True.
-  # 2. If you want to finetune the pretrained models we provide in the docs,
-  #    you should set load_static_weights as False.
-  load_static_weights: True
   cal_metric_during_train: False
   pretrained_model: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
   checkpoints:

configs/det/ch_ppocr_v2.0/ch_det_res18_db_v2.0.yml

@@ -7,11 +7,6 @@ Global:
   save_epoch_step: 1200
   # evaluation is run every 5000 iterations after the 4000th iteration
   eval_batch_step: [3000, 2000]
-  # 1. If pretrained_model is saved in static mode, such as classification pretrained model
-  #    from static branch, load_static_weights must be set as True.
-  # 2. If you want to finetune the pretrained models we provide in the docs,
-  #    you should set load_static_weights as False.
-  load_static_weights: True
   cal_metric_during_train: False
   pretrained_model: ./pretrain_models/ResNet18_vd_pretrained
   checkpoints:

configs/det/det_mv3_db.yml

@@ -7,11 +7,6 @@ Global:
   save_epoch_step: 1200
   # evaluation is run every 2000 iterations
   eval_batch_step: [0, 2000]
-  # 1. If pretrained_model is saved in static mode, such as classification pretrained model
-  #    from static branch, load_static_weights must be set as True.
-  # 2. If you want to finetune the pretrained models we provide in the docs,
-  #    you should set load_static_weights as False.
-  load_static_weights: True
   cal_metric_during_train: False
   pretrained_model: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
   checkpoints:

configs/det/det_mv3_east.yml

@@ -7,11 +7,6 @@ Global:
   save_epoch_step: 1000
   # evaluation is run every 5000 iterations after the 4000th iteration
   eval_batch_step: [4000, 5000]
-  # 1. If pretrained_model is saved in static mode, such as classification pretrained model
-  #    from static branch, load_static_weights must be set as True.
-  # 2. If you want to finetune the pretrained models we provide in the docs,
-  #    you should set load_static_weights as False.
-  load_static_weights: True
   cal_metric_during_train: False
   pretrained_model: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
   checkpoints: 

configs/det/det_r50_vd_db.yml

@@ -7,11 +7,6 @@ Global:
   save_epoch_step: 1200
   # evaluation is run every 2000 iterations
   eval_batch_step: [0,2000]
-  # 1. If pretrained_model is saved in static mode, such as classification pretrained model
-  #    from static branch, load_static_weights must be set as True.
-  # 2. If you want to finetune the pretrained models we provide in the docs,
-  #    you should set load_static_weights as False.
-  load_static_weights: True
   cal_metric_during_train: False
   pretrained_model: ./pretrain_models/ResNet50_vd_ssld_pretrained
   checkpoints:

configs/det/det_r50_vd_east.yml

@@ -7,11 +7,6 @@ Global:
   save_epoch_step: 1000
   # evaluation is run every 5000 iterations after the 4000th iteration
   eval_batch_step: [4000, 5000]
-  # 1. If pretrained_model is saved in static mode, such as classification pretrained model
-  #    from static branch, load_static_weights must be set as True.
-  # 2. If you want to finetune the pretrained models we provide in the docs,
-  #    you should set load_static_weights as False.
-  load_static_weights: True
   cal_metric_during_train: False
   pretrained_model: ./pretrain_models/ResNet50_vd_pretrained/
   checkpoints: 

configs/det/det_r50_vd_sast_icdar15.yml

@@ -7,19 +7,15 @@ Global:
   save_epoch_step: 1000
   # evaluation is run every 5000 iterations after the 4000th iteration
   eval_batch_step: [4000, 5000]
-  # 1. If pretrained_model is saved in static mode, such as classification pretrained model
-  #    from static branch, load_static_weights must be set as True.
-  # 2. If you want to finetune the pretrained models we provide in the docs,
-  #    you should set load_static_weights as False.
-  load_static_weights: True
   cal_metric_during_train: False
   pretrained_model: ./pretrain_models/ResNet50_vd_ssld_pretrained/
-  checkpoints: 
+  checkpoints:
   save_inference_dir:
   use_visualdl: False
-  infer_img: 
+  infer_img:
   save_res_path: ./output/sast_r50_vd_ic15/predicts_sast.txt
 
+
 Architecture:
   model_type: det
   algorithm: SAST

configs/det/det_r50_vd_sast_totaltext.yml

@@ -7,11 +7,6 @@ Global:
   save_epoch_step: 1000
   # evaluation is run every 5000 iterations after the 4000th iteration
   eval_batch_step: [4000, 5000]
-  # 1. If pretrained_model is saved in static mode, such as classification pretrained model
-  #    from static branch, load_static_weights must be set as True.
-  # 2. If you want to finetune the pretrained models we provide in the docs,
-  #    you should set load_static_weights as False.
-  load_static_weights: True
   cal_metric_during_train: False
   pretrained_model: ./pretrain_models/ResNet50_vd_ssld_pretrained/
   checkpoints: 

configs/e2e/e2e_r50_vd_pg.yml

+Global:
+  use_gpu: True
+  epoch_num: 600
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/pgnet_r50_vd_totaltext/
+  save_epoch_step: 10
+  # evaluation is run every 0 iterationss after the 1000th iteration
+  eval_batch_step: [ 0, 1000 ]
+  cal_metric_during_train: False
+  pretrained_model:
+  checkpoints:
+  save_inference_dir:
+  use_visualdl: False
+  infer_img:
+  valid_set: totaltext # two mode: totaltext valid curved words, partvgg valid non-curved words
+  save_res_path: ./output/pgnet_r50_vd_totaltext/predicts_pgnet.txt
+  character_dict_path: ppocr/utils/ic15_dict.txt
+  character_type: EN
+  max_text_length: 50 # the max length in seq
+  max_text_nums: 30 # the max seq nums in a pic
+  tcl_len: 64
+
+Architecture:
+  model_type: e2e
+  algorithm: PGNet
+  Transform:
+  Backbone:
+    name: ResNet
+    layers: 50
+  Neck:
+    name: PGFPN
+  Head:
+    name: PGHead
+
+Loss:
+  name: PGLoss
+  tcl_bs: 64
+  max_text_length: 50 # the same as Global: max_text_length
+  max_text_nums: 30 # the same as Global：max_text_nums
+  pad_num: 36 # the length of dict for pad
+
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+    learning_rate: 0.001
+  regularizer:
+    name: 'L2'
+    factor: 0
+
+
+PostProcess:
+  name: PGPostProcess
+  score_thresh: 0.5
+  mode: fast   # fast or slow two ways
+
+Metric:
+  name: E2EMetric
+  mode: A   # two ways for eval, A: label from txt,  B: label from gt_mat
+  gt_mat_dir:  ./train_data/total_text/gt  # the dir of gt_mat
+  character_dict_path: ppocr/utils/ic15_dict.txt
+  main_indicator: f_score_e2e
+
+Train:
+  dataset:
+    name: PGDataSet
+    data_dir: ./train_data/total_text/train
+    label_file_list: [./train_data/total_text/train/train.txt]
+    ratio_list: [1.0]
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - E2ELabelEncodeTrain:
+      - PGProcessTrain:
+          batch_size: 14  # same as loader: batch_size_per_card
+          min_crop_size: 24
+          min_text_size: 4
+          max_text_size: 512
+      - KeepKeys:
+          keep_keys: [ 'images', 'tcl_maps', 'tcl_label_maps', 'border_maps','direction_maps', 'training_masks', 'label_list', 'pos_list', 'pos_mask' ] # dataloader will return list in this order
+  loader:
+    shuffle: True
+    drop_last: True
+    batch_size_per_card: 14
+    num_workers: 16
+
+Eval:
+  dataset:
+    name: PGDataSet
+    data_dir: ./train_data/total_text/test
+    label_file_list: [./train_data/total_text/test/test.txt]
+    transforms:
+      - DecodeImage: # load image
+          img_mode: RGB
+          channel_first: False
+      - E2ELabelEncodeTest:
+      - E2EResizeForTest:
+          max_side_len: 768
+      - NormalizeImage:
+          scale: 1./255.
+          mean: [ 0.485, 0.456, 0.406 ]
+          std: [ 0.229, 0.224, 0.225 ]
+          order: 'hwc'
+      - ToCHWImage:
+      - KeepKeys:
+          keep_keys: [ 'image', 'shape', 'polys', 'texts', 'ignore_tags', 'img_id']
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 1 # must be 1
+    num_workers: 2
\ No newline at end of file

configs/rec/ch_ppocr_v2.0/rec_chinese_common_train_v2.0.yml

@@ -19,6 +19,7 @@ Global:
   max_text_length: 25
   infer_mode: False
   use_space_char: True
+  save_res_path: ./output/rec/predicts_chinese_common_v2.0.txt
 
 
 Optimizer:

configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml

@@ -19,6 +19,7 @@ Global:
   max_text_length: 25
   infer_mode: False
   use_space_char: True
+  save_res_path: ./output/rec/predicts_chinese_lite_v2.0.txt
 
 
 Optimizer:

configs/rec/multi_language/generate_multi_language_configs.py

@@ -19,21 +19,56 @@ import logging
 logging.basicConfig(level=logging.INFO)
 
 support_list = {
-    'it':'italian', 'xi':'spanish', 'pu':'portuguese', 'ru':'russian', 'ar':'arabic',
-    'ta':'tamil', 'ug':'uyghur', 'fa':'persian', 'ur':'urdu', 'rs':'serbian latin',
-    'oc':'occitan', 'rsc':'serbian cyrillic', 'bg':'bulgarian', 'uk':'ukranian', 'be':'belarusian',
-    'te':'telugu', 'ka':'kannada', 'chinese_cht':'chinese tradition','hi':'hindi','mr':'marathi',
-    'ne':'nepali',
+    'it': 'italian',
+    'xi': 'spanish',
+    'pu': 'portuguese',
+    'ru': 'russian',
+    'ar': 'arabic',
+    'ta': 'tamil',
+    'ug': 'uyghur',
+    'fa': 'persian',
+    'ur': 'urdu',
+    'rs': 'serbian latin',
+    'oc': 'occitan',
+    'rsc': 'serbian cyrillic',
+    'bg': 'bulgarian',
+    'uk': 'ukranian',
+    'be': 'belarusian',
+    'te': 'telugu',
+    'ka': 'kannada',
+    'chinese_cht': 'chinese tradition',
+    'hi': 'hindi',
+    'mr': 'marathi',
+    'ne': 'nepali',
 }
-assert(
-    os.path.isfile("./rec_multi_language_lite_train.yml")
-    ),"Loss basic configuration file rec_multi_language_lite_train.yml.\
+
+latin_lang = [
+    'af', 'az', 'bs', 'cs', 'cy', 'da', 'de', 'es', 'et', 'fr', 'ga', 'hr',
+    'hu', 'id', 'is', 'it', 'ku', 'la', 'lt', 'lv', 'mi', 'ms', 'mt', 'nl',
+    'no', 'oc', 'pi', 'pl', 'pt', 'ro', 'rs_latin', 'sk', 'sl', 'sq', 'sv',
+    'sw', 'tl', 'tr', 'uz', 'vi', 'latin'
+]
+arabic_lang = ['ar', 'fa', 'ug', 'ur']
+cyrillic_lang = [
+    'ru', 'rs_cyrillic', 'be', 'bg', 'uk', 'mn', 'abq', 'ady', 'kbd', 'ava',
+    'dar', 'inh', 'che', 'lbe', 'lez', 'tab', 'cyrillic'
+]
+devanagari_lang = [
+    'hi', 'mr', 'ne', 'bh', 'mai', 'ang', 'bho', 'mah', 'sck', 'new', 'gom',
+    'sa', 'bgc', 'devanagari'
+]
+multi_lang = latin_lang + arabic_lang + cyrillic_lang + devanagari_lang
+
+assert (os.path.isfile("./rec_multi_language_lite_train.yml")
+        ), "Loss basic configuration file rec_multi_language_lite_train.yml.\
 You can download it from \
 https://github.com/PaddlePaddle/PaddleOCR/tree/dygraph/configs/rec/multi_language/"
- 
-global_config = yaml.load(open("./rec_multi_language_lite_train.yml", 'rb'), Loader=yaml.Loader)
+
+global_config = yaml.load(
+    open("./rec_multi_language_lite_train.yml", 'rb'), Loader=yaml.Loader)
 project_path = os.path.abspath(os.path.join(os.getcwd(), "../../../"))
 
+
 class ArgsParser(ArgumentParser):
     def __init__(self):
         super(ArgsParser, self).__init__(
@@ -41,15 +76,30 @@ class ArgsParser(ArgumentParser):
         self.add_argument(
             "-o", "--opt", nargs='+', help="set configuration options")
         self.add_argument(
-            "-l", "--language", nargs='+', help="set language type, support {}".format(support_list))
+            "-l",
+            "--language",
+            nargs='+',
+            help="set language type, support {}".format(support_list))
         self.add_argument(
-            "--train",type=str,help="you can use this command to change the train dataset default path")
+            "--train",
+            type=str,
+            help="you can use this command to change the train dataset default path"
+        )
         self.add_argument(
-            "--val",type=str,help="you can use this command to change the eval dataset default path")
+            "--val",
+            type=str,
+            help="you can use this command to change the eval dataset default path"
+        )
         self.add_argument(
-            "--dict",type=str,help="you can use this command to change the dictionary default path")
+            "--dict",
+            type=str,
+            help="you can use this command to change the dictionary default path"
+        )
         self.add_argument(
-            "--data_dir",type=str,help="you can use this command to change the dataset default root path")
+            "--data_dir",
+            type=str,
+            help="you can use this command to change the dataset default root path"
+        )
 
     def parse_args(self, argv=None):
         args = super(ArgsParser, self).parse_args(argv)
@@ -68,21 +118,37 @@ class ArgsParser(ArgumentParser):
         return config
 
     def _set_language(self, type):
-        assert(type),"please use -l or --language to choose language type"
+        lang = type[0]
+        assert (type), "please use -l or --language to choose language type"
         assert(
-                type[0] in support_list.keys()
+                lang in support_list.keys() or lang in multi_lang
                ),"the sub_keys(-l or --language) can only be one of support list: \n{},\nbut get: {}, " \
-                 "please check your running command".format(support_list, type)
-        global_config['Global']['character_dict_path'] = 'ppocr/utils/dict/{}_dict.txt'.format(type[0])
-        global_config['Global']['save_model_dir'] = './output/rec_{}_lite'.format(type[0])
-        global_config['Train']['dataset']['label_file_list'] = ["train_data/{}_train.txt".format(type[0])]
-        global_config['Eval']['dataset']['label_file_list'] = ["train_data/{}_val.txt".format(type[0])]
-        global_config['Global']['character_type'] = type[0]
-        assert(
-                os.path.isfile(os.path.join(project_path,global_config['Global']['character_dict_path']))
-              ),"Loss default dictionary file {}_dict.txt.You can download it from \
-https://github.com/PaddlePaddle/PaddleOCR/tree/dygraph/ppocr/utils/dict/".format(type[0])
-        return type[0]
+                 "please check your running command".format(multi_lang, type)
+        if lang in latin_lang:
+            lang = "latin"
+        elif lang in arabic_lang:
+            lang = "arabic"
+        elif lang in cyrillic_lang:
+            lang = "cyrillic"
+        elif lang in devanagari_lang:
+            lang = "devanagari"
+        global_config['Global'][
+            'character_dict_path'] = 'ppocr/utils/dict/{}_dict.txt'.format(lang)
+        global_config['Global'][
+            'save_model_dir'] = './output/rec_{}_lite'.format(lang)
+        global_config['Train']['dataset'][
+            'label_file_list'] = ["train_data/{}_train.txt".format(lang)]
+        global_config['Eval']['dataset'][
+            'label_file_list'] = ["train_data/{}_val.txt".format(lang)]
+        global_config['Global']['character_type'] = lang
+        assert (
+            os.path.isfile(
+                os.path.join(project_path, global_config['Global'][
+                    'character_dict_path']))
+        ), "Loss default dictionary file {}_dict.txt.You can download it from \
+https://github.com/PaddlePaddle/PaddleOCR/tree/dygraph/ppocr/utils/dict/".format(
+            lang)
+        return lang
 
 
 def merge_config(config):
@@ -110,43 +176,51 @@ def merge_config(config):
                     cur[sub_key] = value
                 else:
                     cur = cur[sub_key]
-                    
+
+
 def loss_file(path):
-    assert(
-            os.path.exists(path)
-          ),"There is no such file:{},Please do not forget to put in the specified file".format(path)
+    assert (
+        os.path.exists(path)
+    ), "There is no such file:{},Please do not forget to put in the specified file".format(
+        path)
+
 
-        
 if __name__ == '__main__':
     FLAGS = ArgsParser().parse_args()
     merge_config(FLAGS.opt)
     save_file_path = 'rec_{}_lite_train.yml'.format(FLAGS.language)
     if os.path.isfile(save_file_path):
         os.remove(save_file_path)
-        
+
     if FLAGS.train:
         global_config['Train']['dataset']['label_file_list'] = [FLAGS.train]
-        train_label_path = os.path.join(project_path,FLAGS.train)
+        train_label_path = os.path.join(project_path, FLAGS.train)
         loss_file(train_label_path)
     if FLAGS.val:
         global_config['Eval']['dataset']['label_file_list'] = [FLAGS.val]
-        eval_label_path = os.path.join(project_path,FLAGS.val)
+        eval_label_path = os.path.join(project_path, FLAGS.val)
         loss_file(eval_label_path)
     if FLAGS.dict:
         global_config['Global']['character_dict_path'] = FLAGS.dict
-        dict_path = os.path.join(project_path,FLAGS.dict)
+        dict_path = os.path.join(project_path, FLAGS.dict)
         loss_file(dict_path)
     if FLAGS.data_dir:
         global_config['Eval']['dataset']['data_dir'] = FLAGS.data_dir
         global_config['Train']['dataset']['data_dir'] = FLAGS.data_dir
-        data_dir = os.path.join(project_path,FLAGS.data_dir)
+        data_dir = os.path.join(project_path, FLAGS.data_dir)
         loss_file(data_dir)
-        
+
     with open(save_file_path, 'w') as f:
-        yaml.dump(dict(global_config), f, default_flow_style=False, sort_keys=False)
+        yaml.dump(
+            dict(global_config), f, default_flow_style=False, sort_keys=False)
     logging.info("Project path is          :{}".format(project_path))
-    logging.info("Train list path set to   :{}".format(global_config['Train']['dataset']['label_file_list'][0]))
-    logging.info("Eval list path set to    :{}".format(global_config['Eval']['dataset']['label_file_list'][0]))
-    logging.info("Dataset root path set to :{}".format(global_config['Eval']['dataset']['data_dir']))
-    logging.info("Dict path set to         :{}".format(global_config['Global']['character_dict_path']))
-    logging.info("Config file set to       :configs/rec/multi_language/{}".format(save_file_path))
+    logging.info("Train list path set to   :{}".format(global_config['Train'][
+        'dataset']['label_file_list'][0]))
+    logging.info("Eval list path set to    :{}".format(global_config['Eval'][
+        'dataset']['label_file_list'][0]))
+    logging.info("Dataset root path set to :{}".format(global_config['Eval'][
+        'dataset']['data_dir']))
+    logging.info("Dict path set to         :{}".format(global_config['Global'][
+        'character_dict_path']))
+    logging.info("Config file set to       :configs/rec/multi_language/{}".
+                 format(save_file_path))

configs/rec/multi_language/rec_arabic_lite_train.yml

+Global:
+  use_gpu: true
+  epoch_num: 500
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec_arabic_lite
+  save_epoch_step: 3
+  eval_batch_step:
+  - 0
+  - 2000
+  cal_metric_during_train: true
+  pretrained_model: null
+  checkpoints: null
+  save_inference_dir: null
+  use_visualdl: false
+  infer_img: null
+  character_dict_path: ppocr/utils/dict/arabic_dict.txt
+  character_type: arabic
+  max_text_length: 25
+  infer_mode: false
+  use_space_char: true
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+    name: Cosine
+    learning_rate: 0.001
+  regularizer:
+    name: L2
+    factor: 1.0e-05
+Architecture:
+  model_type: rec
+  algorithm: CRNN
+  Transform: null
+  Backbone:
+    name: MobileNetV3
+    scale: 0.5
+    model_name: small
+    small_stride:
+    - 1
+    - 2
+    - 2
+    - 2
+  Neck:
+    name: SequenceEncoder
+    encoder_type: rnn
+    hidden_size: 48
+  Head:
+    name: CTCHead
+    fc_decay: 1.0e-05
+Loss:
+  name: CTCLoss
+PostProcess:
+  name: CTCLabelDecode
+Metric:
+  name: RecMetric
+  main_indicator: acc
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: train_data/
+    label_file_list:
+    - train_data/arabic_train.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - RecAug: null
+    - CTCLabelEncode: null
+    - RecResizeImg:
+        image_shape:
+        - 3
+        - 32
+        - 320
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: true
+    batch_size_per_card: 256
+    drop_last: true
+    num_workers: 8
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: train_data/
+    label_file_list:
+    - train_data/arabic_val.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - CTCLabelEncode: null
+    - RecResizeImg:
+        image_shape:
+        - 3
+        - 32
+        - 320
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: false
+    drop_last: false
+    batch_size_per_card: 256
+    num_workers: 8

configs/rec/multi_language/rec_cyrillic_lite_train.yml

+Global:
+  use_gpu: true
+  epoch_num: 500
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec_cyrillic_lite
+  save_epoch_step: 3
+  eval_batch_step:
+  - 0
+  - 2000
+  cal_metric_during_train: true
+  pretrained_model: null
+  checkpoints: null
+  save_inference_dir: null
+  use_visualdl: false
+  infer_img: null
+  character_dict_path: ppocr/utils/dict/cyrillic_dict.txt
+  character_type: cyrillic
+  max_text_length: 25
+  infer_mode: false
+  use_space_char: true
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+    name: Cosine
+    learning_rate: 0.001
+  regularizer:
+    name: L2
+    factor: 1.0e-05
+Architecture:
+  model_type: rec
+  algorithm: CRNN
+  Transform: null
+  Backbone:
+    name: MobileNetV3
+    scale: 0.5
+    model_name: small
+    small_stride:
+    - 1
+    - 2
+    - 2
+    - 2
+  Neck:
+    name: SequenceEncoder
+    encoder_type: rnn
+    hidden_size: 48
+  Head:
+    name: CTCHead
+    fc_decay: 1.0e-05
+Loss:
+  name: CTCLoss
+PostProcess:
+  name: CTCLabelDecode
+Metric:
+  name: RecMetric
+  main_indicator: acc
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: train_data/
+    label_file_list:
+    - train_data/cyrillic_train.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - RecAug: null
+    - CTCLabelEncode: null
+    - RecResizeImg:
+        image_shape:
+        - 3
+        - 32
+        - 320
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: true
+    batch_size_per_card: 256
+    drop_last: true
+    num_workers: 8
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: train_data/
+    label_file_list:
+    - train_data/cyrillic_val.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - CTCLabelEncode: null
+    - RecResizeImg:
+        image_shape:
+        - 3
+        - 32
+        - 320
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: false
+    drop_last: false
+    batch_size_per_card: 256
+    num_workers: 8

configs/rec/multi_language/rec_devanagari_lite_train.yml

+Global:
+  use_gpu: true
+  epoch_num: 500
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec_devanagari_lite
+  save_epoch_step: 3
+  eval_batch_step:
+  - 0
+  - 2000
+  cal_metric_during_train: true
+  pretrained_model: null
+  checkpoints: null
+  save_inference_dir: null
+  use_visualdl: false
+  infer_img: null
+  character_dict_path: ppocr/utils/dict/devanagari_dict.txt
+  character_type: devanagari
+  max_text_length: 25
+  infer_mode: false
+  use_space_char: true
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+    name: Cosine
+    learning_rate: 0.001
+  regularizer:
+    name: L2
+    factor: 1.0e-05
+Architecture:
+  model_type: rec
+  algorithm: CRNN
+  Transform: null
+  Backbone:
+    name: MobileNetV3
+    scale: 0.5
+    model_name: small
+    small_stride:
+    - 1
+    - 2
+    - 2
+    - 2
+  Neck:
+    name: SequenceEncoder
+    encoder_type: rnn
+    hidden_size: 48
+  Head:
+    name: CTCHead
+    fc_decay: 1.0e-05
+Loss:
+  name: CTCLoss
+PostProcess:
+  name: CTCLabelDecode
+Metric:
+  name: RecMetric
+  main_indicator: acc
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: train_data/
+    label_file_list:
+    - train_data/devanagari_train.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - RecAug: null
+    - CTCLabelEncode: null
+    - RecResizeImg:
+        image_shape:
+        - 3
+        - 32
+        - 320
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: true
+    batch_size_per_card: 256
+    drop_last: true
+    num_workers: 8
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: train_data/
+    label_file_list:
+    - train_data/devanagari_val.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - CTCLabelEncode: null
+    - RecResizeImg:
+        image_shape:
+        - 3
+        - 32
+        - 320
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: false
+    drop_last: false
+    batch_size_per_card: 256
+    num_workers: 8

configs/rec/multi_language/rec_en_number_lite_train.yml

@@ -15,11 +15,11 @@ Global:
   use_visualdl: False
   infer_img:
   # for data or label process
-  character_dict_path: ppocr/utils/dict/en_dict.txt
+  character_dict_path: ppocr/utils/en_dict.txt
   character_type: EN
   max_text_length: 25
   infer_mode: False
-  use_space_char: False
+  use_space_char: True
 
 
 Optimizer:

configs/rec/multi_language/rec_latin_lite_train.yml

+Global:
+  use_gpu: true
+  epoch_num: 500
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec_latin_lite
+  save_epoch_step: 3
+  eval_batch_step:
+  - 0
+  - 2000
+  cal_metric_during_train: true
+  pretrained_model: null
+  checkpoints: null
+  save_inference_dir: null
+  use_visualdl: false
+  infer_img: null
+  character_dict_path: ppocr/utils/dict/latin_dict.txt
+  character_type: latin
+  max_text_length: 25
+  infer_mode: false
+  use_space_char: true
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+    name: Cosine
+    learning_rate: 0.001
+  regularizer:
+    name: L2
+    factor: 1.0e-05
+Architecture:
+  model_type: rec
+  algorithm: CRNN
+  Transform: null
+  Backbone:
+    name: MobileNetV3
+    scale: 0.5
+    model_name: small
+    small_stride:
+    - 1
+    - 2
+    - 2
+    - 2
+  Neck:
+    name: SequenceEncoder
+    encoder_type: rnn
+    hidden_size: 48
+  Head:
+    name: CTCHead
+    fc_decay: 1.0e-05
+Loss:
+  name: CTCLoss
+PostProcess:
+  name: CTCLabelDecode
+Metric:
+  name: RecMetric
+  main_indicator: acc
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: train_data/
+    label_file_list:
+    - train_data/latin_train.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - RecAug: null
+    - CTCLabelEncode: null
+    - RecResizeImg:
+        image_shape:
+        - 3
+        - 32
+        - 320
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: true
+    batch_size_per_card: 256
+    drop_last: true
+    num_workers: 8
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: train_data/
+    label_file_list:
+    - train_data/latin_val.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - CTCLabelEncode: null
+    - RecResizeImg:
+        image_shape:
+        - 3
+        - 32
+        - 320
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: false
+    drop_last: false
+    batch_size_per_card: 256
+    num_workers: 8

configs/rec/rec_icdar15_train.yml

@@ -19,6 +19,7 @@ Global:
   max_text_length: 25
   infer_mode: False
   use_space_char: False
+  save_res_path: ./output/rec/predicts_ic15.txt
 
 Optimizer:
   name: Adam
@@ -81,7 +82,7 @@ Eval:
   dataset:
     name: SimpleDataSet
     data_dir: ./train_data/
-    label_file_list: ["./train_data/train_list.txt"]
+    label_file_list: ["./train_data/val_list.txt"]
     transforms:
       - DecodeImage: # load image
           img_mode: BGR

configs/rec/rec_mv3_none_bilstm_ctc.yml

@@ -19,6 +19,7 @@ Global:
   max_text_length: 25
   infer_mode: False
   use_space_char: False
+  save_res_path: ./output/rec/predicts_mv3_none_bilstm_ctc.txt
 
 Optimizer:
   name: Adam

configs/rec/rec_mv3_none_none_ctc.yml

@@ -19,6 +19,7 @@ Global:
   max_text_length: 25
   infer_mode: False
   use_space_char: False
+  save_res_path: ./output/rec/predicts_mv3_none_none_ctc.txt
 
 Optimizer:
   name: Adam

configs/rec/rec_mv3_tps_bilstm_att.yml

@@ -19,6 +19,7 @@ Global:
   max_text_length: 25
   infer_mode: False
   use_space_char: False
+  save_res_path: ./output/rec/predicts_mv3_tps_bilstm_att.txt
 
 
 Optimizer:

configs/rec/rec_mv3_tps_bilstm_ctc.yml

@@ -19,6 +19,7 @@ Global:
   max_text_length: 25
   infer_mode: False
   use_space_char: False
+  save_res_path: ./output/rec/predicts_mv3_tps_bilstm_ctc.txt
 
 Optimizer:
   name: Adam

configs/rec/rec_r34_vd_none_bilstm_ctc.yml

@@ -19,6 +19,7 @@ Global:
   max_text_length: 25
   infer_mode: False
   use_space_char: False
+  save_res_path: ./output/rec/predicts_r34_vd_none_bilstm_ctc.txt
 
 Optimizer:
   name: Adam

configs/rec/rec_r34_vd_none_none_ctc.yml

@@ -19,6 +19,7 @@ Global:
   max_text_length: 25
   infer_mode: False
   use_space_char: False
+  save_res_path: ./output/rec/predicts_r34_vd_none_none_ctc.txt
 
 Optimizer:
   name: Adam

configs/rec/rec_r34_vd_tps_bilstm_att.yml

@@ -19,6 +19,7 @@ Global:
   max_text_length: 25
   infer_mode: False
   use_space_char: False
+  save_res_path: ./output/rec/predicts_b3_rare_r34_none_gru.txt
 
 
 Optimizer:

configs/rec/rec_r34_vd_tps_bilstm_ctc.yml

@@ -19,6 +19,7 @@ Global:
   max_text_length: 25
   infer_mode: False
   use_space_char: False
+  save_res_path: ./output/rec/predicts_r34_vd_tps_bilstm_ctc.txt
 
 Optimizer:
   name: Adam
@@ -37,7 +38,7 @@ Architecture:
     name: TPS
     num_fiducial: 20
     loc_lr: 0.1
-    model_name: small
+    model_name: large
   Backbone:
     name: ResNet
     layers: 34

configs/rec/rec_r50_fpn_srn.yml

@@ -20,6 +20,7 @@ Global:
   num_heads: 8
   infer_mode: False
   use_space_char: False
+  save_res_path: ./output/rec/predicts_srn.txt
 
 
 Optimizer:

deploy/android_demo/.gitignore

+*.iml
+.gradle
+/local.properties
+/.idea/*
+.DS_Store
+/build
+/captures
+.externalNativeBuild
+

deploy/android_demo/README.md

+# 如何快速测试
+### 1. 安装最新版本的Android Studio
+可以从 https://developer.android.com/studio 下载。本Demo使用是4.0版本Android Studio编写。
+
+### 2. 按照NDK 20 以上版本
+Demo测试的时候使用的是NDK 20b版本，20版本以上均可以支持编译成功。
+
+如果您是初学者，可以用以下方式安装和测试NDK编译环境。
+点击 File -> New ->New Project，  新建  "Native C++" project
+
+### 3. 导入项目
+点击 File->New->Import Project...， 然后跟着Android Studio的引导导入
+
+
+# 获得更多支持
+前往[端计算模型生成平台EasyEdge](https://ai.baidu.com/easyedge/app/open_source_demo?referrerUrl=paddlelite)，获得更多开发支持：
+
+- Demo APP：可使用手机扫码安装，方便手机端快速体验文字识别
+- SDK：模型被封装为适配不同芯片硬件和操作系统SDK，包括完善的接口，方便进行二次开发

deploy/android_demo/app/.gitignore

+/build

deploy/android_demo/app/build.gradle

+import java.security.MessageDigest
+
+apply plugin: 'com.android.application'
+
+android {
+    compileSdkVersion 29
+    defaultConfig {
+        applicationId "com.baidu.paddle.lite.demo.ocr"
+        minSdkVersion 23
+        targetSdkVersion 29
+        versionCode 1
+        versionName "1.0"
+        testInstrumentationRunner "android.support.test.runner.AndroidJUnitRunner"
+        externalNativeBuild {
+            cmake {
+                cppFlags "-std=c++11 -frtti -fexceptions -Wno-format"
+                arguments '-DANDROID_PLATFORM=android-23', '-DANDROID_STL=c++_shared' ,"-DANDROID_ARM_NEON=TRUE"
+            }
+        }
+        ndk {
+            // abiFilters "arm64-v8a", "armeabi-v7a"
+            abiFilters   "arm64-v8a", "armeabi-v7a"
+            ldLibs "jnigraphics"
+        }
+    }
+    buildTypes {
+        release {
+            minifyEnabled false
+            proguardFiles getDefaultProguardFile('proguard-android-optimize.txt'), 'proguard-rules.pro'
+        }
+    }
+    externalNativeBuild {
+        cmake {
+            path "src/main/cpp/CMakeLists.txt"
+            version "3.10.2"
+        }
+    }
+}
+
+dependencies {
+    implementation fileTree(include: ['*.jar'], dir: 'libs')
+    implementation 'androidx.appcompat:appcompat:1.1.0'
+    implementation 'androidx.constraintlayout:constraintlayout:1.1.3'
+    testImplementation 'junit:junit:4.12'
+    androidTestImplementation 'com.android.support.test:runner:1.0.2'
+    androidTestImplementation 'com.android.support.test.espresso:espresso-core:3.0.2'
+}
+
+def archives = [
+        [
+                'src' : 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/paddle_lite_libs_v2_9_0.tar.gz',
+                'dest': 'PaddleLite'
+        ],
+        [
+                'src' : 'https://paddlelite-demo.bj.bcebos.com/libs/android/opencv-4.2.0-android-sdk.tar.gz',
+                'dest': 'OpenCV'
+        ],
+        [
+                'src' : 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ocr_v2_for_cpu.tar.gz',
+                'dest' : 'src/main/assets/models'
+        ],
+        [
+                'src' : 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/lite/ch_dict.tar.gz',
+                'dest' : 'src/main/assets/labels'
+        ]
+]
+
+task downloadAndExtractArchives(type: DefaultTask) {
+    doFirst {
+        println "Downloading and extracting archives including libs and models"
+    }
+    doLast {
+        // Prepare cache folder for archives
+        String cachePath = "cache"
+        if (!file("${cachePath}").exists()) {
+            mkdir "${cachePath}"
+        }
+        archives.eachWithIndex { archive, index ->
+            MessageDigest messageDigest = MessageDigest.getInstance('MD5')
+            messageDigest.update(archive.src.bytes)
+            String cacheName = new BigInteger(1, messageDigest.digest()).toString(32)
+            // Download the target archive if not exists
+            boolean copyFiles = !file("${archive.dest}").exists()
+            if (!file("${cachePath}/${cacheName}.tar.gz").exists()) {
+                ant.get(src: archive.src, dest: file("${cachePath}/${cacheName}.tar.gz"))
+                copyFiles = true; // force to copy files from the latest archive files
+            }
+            // Extract the target archive if its dest path does not exists
+            if (copyFiles) {
+                copy {
+                    from tarTree("${cachePath}/${cacheName}.tar.gz")
+                    into "${archive.dest}"
+                }
+            }
+        }
+    }
+}
+preBuild.dependsOn downloadAndExtractArchives
\ No newline at end of file

deploy/android_demo/app/proguard-rules.pro

+# Add project specific ProGuard rules here.
+# You can control the set of applied configuration files using the
+# proguardFiles setting in build.gradle.
+#
+# For more details, see
+#   http://developer.android.com/guide/developing/tools/proguard.html
+
+# If your project uses WebView with JS, uncomment the following
+# and specify the fully qualified class name to the JavaScript interface
+# class:
+#-keepclassmembers class fqcn.of.javascript.interface.for.webview {
+#   public *;
+#}
+
+# Uncomment this to preserve the line number information for
+# debugging stack traces.
+#-keepattributes SourceFile,LineNumberTable
+
+# If you keep the line number information, uncomment this to
+# hide the original source file name.
+#-renamesourcefileattribute SourceFile

deploy/android_demo/app/src/androidTest/java/com/baidu/paddle/lite/demo/ocr/ExampleInstrumentedTest.java

+package com.baidu.paddle.lite.demo.ocr;
+
+import android.content.Context;
+import android.support.test.InstrumentationRegistry;
+import android.support.test.runner.AndroidJUnit4;
+
+import org.junit.Test;
+import org.junit.runner.RunWith;
+
+import static org.junit.Assert.*;
+
+/**
+ * Instrumented test, which will execute on an Android device.
+ *
+ * @see <a href="http://d.android.com/tools/testing">Testing documentation</a>
+ */
+@RunWith(AndroidJUnit4.class)
+public class ExampleInstrumentedTest {
+    @Test
+    public void useAppContext() {
+        // Context of the app under test.
+        Context appContext = InstrumentationRegistry.getTargetContext();
+
+        assertEquals("com.baidu.paddle.lite.demo", appContext.getPackageName());
+    }
+}

deploy/android_demo/app/src/main/AndroidManifest.xml

+<?xml version="1.0" encoding="utf-8"?>
+<manifest xmlns:android="http://schemas.android.com/apk/res/android"
+          package="com.baidu.paddle.lite.demo.ocr">
+
+    <uses-permission android:name="android.permission.WRITE_EXTERNAL_STORAGE"/>
+    <uses-permission android:name="android.permission.READ_EXTERNAL_STORAGE"/>
+    <uses-permission android:name="android.permission.CAMERA"/>
+
+
+    <application
+            android:allowBackup="true"
+            android:icon="@mipmap/ic_launcher"
+            android:label="@string/app_name"
+            android:roundIcon="@mipmap/ic_launcher_round"
+            android:supportsRtl="true"
+            android:theme="@style/AppTheme">
+        <!-- to test MiniActivity, change this to com.baidu.paddle.lite.demo.ocr.MiniActivity -->
+        <activity android:name="com.baidu.paddle.lite.demo.ocr.MainActivity">
+            <intent-filter>
+                <action android:name="android.intent.action.MAIN"/>
+                <category android:name="android.intent.category.LAUNCHER"/>
+            </intent-filter>
+        </activity>
+        <activity
+                android:name="com.baidu.paddle.lite.demo.ocr.SettingsActivity"
+                android:label="Settings">
+        </activity>
+        <provider
+            android:name="androidx.core.content.FileProvider"
+            android:authorities="com.baidu.paddle.lite.demo.ocr.fileprovider"
+            android:exported="false"
+            android:grantUriPermissions="true">
+            <meta-data
+                android:name="android.support.FILE_PROVIDER_PATHS"
+                android:resource="@xml/file_paths"></meta-data>
+        </provider>
+    </application>
+
+</manifest>
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/CMakeLists.txt

+# For more information about using CMake with Android Studio, read the
+# documentation: https://d.android.com/studio/projects/add-native-code.html
+
+# Sets the minimum version of CMake required to build the native library.
+
+cmake_minimum_required(VERSION 3.4.1)
+
+# Creates and names a library, sets it as either STATIC or SHARED, and provides
+# the relative paths to its source code. You can define multiple libraries, and
+# CMake builds them for you. Gradle automatically packages shared libraries with
+# your APK.
+
+set(PaddleLite_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../../../PaddleLite")
+include_directories(${PaddleLite_DIR}/cxx/include)
+
+set(OpenCV_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../../../OpenCV/sdk/native/jni")
+message(STATUS "opencv dir: ${OpenCV_DIR}")
+find_package(OpenCV REQUIRED)
+message(STATUS "OpenCV libraries: ${OpenCV_LIBS}")
+include_directories(${OpenCV_INCLUDE_DIRS})
+aux_source_directory(. SOURCES)
+set(CMAKE_CXX_FLAGS
+        "${CMAKE_CXX_FLAGS} -ffast-math -Ofast -Os"
+        )
+set(CMAKE_CXX_FLAGS
+        "${CMAKE_CXX_FLAGS} -fvisibility=hidden -fvisibility-inlines-hidden -fdata-sections -ffunction-sections"
+        )
+set(CMAKE_SHARED_LINKER_FLAGS
+        "${CMAKE_SHARED_LINKER_FLAGS} -Wl,--gc-sections -Wl,-z,nocopyreloc")
+
+add_library(
+        # Sets the name of the library.
+        Native
+        # Sets the library as a shared library.
+        SHARED
+        # Provides a relative path to your source file(s).
+        ${SOURCES})
+
+find_library(
+        # Sets the name of the path variable.
+        log-lib
+        # Specifies the name of the NDK library that you want CMake to locate.
+        log)
+
+add_library(
+        # Sets the name of the library.
+        paddle_light_api_shared
+        # Sets the library as a shared library.
+        SHARED
+        # Provides a relative path to your source file(s).
+        IMPORTED)
+
+set_target_properties(
+        # Specifies the target library.
+        paddle_light_api_shared
+        # Specifies the parameter you want to define.
+        PROPERTIES
+        IMPORTED_LOCATION
+        ${PaddleLite_DIR}/cxx/libs/${ANDROID_ABI}/libpaddle_light_api_shared.so
+        # Provides the path to the library you want to import.
+)
+
+
+# Specifies libraries CMake should link to your target library. You can link
+# multiple libraries, such as libraries you define in this build script,
+# prebuilt third-party libraries, or system libraries.
+
+target_link_libraries(
+        # Specifies the target library.
+        Native
+        paddle_light_api_shared
+        ${OpenCV_LIBS}
+        GLESv2
+        EGL
+        jnigraphics
+        ${log-lib}
+)
+
+add_custom_command(
+        TARGET Native
+        POST_BUILD
+        COMMAND
+        ${CMAKE_COMMAND} -E copy
+        ${PaddleLite_DIR}/cxx/libs/${ANDROID_ABI}/libc++_shared.so
+        ${CMAKE_LIBRARY_OUTPUT_DIRECTORY}/libc++_shared.so)
+
+add_custom_command(
+        TARGET Native
+        POST_BUILD
+        COMMAND
+        ${CMAKE_COMMAND} -E copy
+        ${PaddleLite_DIR}/cxx/libs/${ANDROID_ABI}/libpaddle_light_api_shared.so
+        ${CMAKE_LIBRARY_OUTPUT_DIRECTORY}/libpaddle_light_api_shared.so)
+
+add_custom_command(
+        TARGET Native
+        POST_BUILD
+        COMMAND
+        ${CMAKE_COMMAND} -E copy
+        ${PaddleLite_DIR}/cxx/libs/${ANDROID_ABI}/libhiai.so
+        ${CMAKE_LIBRARY_OUTPUT_DIRECTORY}/libhiai.so)
+
+add_custom_command(
+        TARGET Native
+        POST_BUILD
+        COMMAND
+        ${CMAKE_COMMAND} -E copy
+        ${PaddleLite_DIR}/cxx/libs/${ANDROID_ABI}/libhiai_ir.so
+        ${CMAKE_LIBRARY_OUTPUT_DIRECTORY}/libhiai_ir.so)
+
+add_custom_command(
+        TARGET Native
+        POST_BUILD
+        COMMAND
+        ${CMAKE_COMMAND} -E copy
+        ${PaddleLite_DIR}/cxx/libs/${ANDROID_ABI}/libhiai_ir_build.so
+        ${CMAKE_LIBRARY_OUTPUT_DIRECTORY}/libhiai_ir_build.so)
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/common.h

+//
+// Created by fu on 4/25/18.
+//
+
+#pragma once
+#import <numeric>
+#import <vector>
+
+#ifdef __ANDROID__
+
+#include <android/log.h>
+
+#define LOG_TAG "OCR_NDK"
+
+#define LOGI(...) __android_log_print(ANDROID_LOG_INFO, LOG_TAG, __VA_ARGS__)
+#define LOGW(...) __android_log_print(ANDROID_LOG_WARN, LOG_TAG, __VA_ARGS__)
+#define LOGE(...) __android_log_print(ANDROID_LOG_ERROR, LOG_TAG, __VA_ARGS__)
+#else
+#include <stdio.h>
+#define LOGI(format, ...)                                                      \
+  fprintf(stdout, "[" LOG_TAG "]" format "\n", ##__VA_ARGS__)
+#define LOGW(format, ...)                                                      \
+  fprintf(stdout, "[" LOG_TAG "]" format "\n", ##__VA_ARGS__)
+#define LOGE(format, ...)                                                      \
+  fprintf(stderr, "[" LOG_TAG "]Error: " format "\n", ##__VA_ARGS__)
+#endif
+
+enum RETURN_CODE { RETURN_OK = 0 };
+
+enum NET_TYPE { NET_OCR = 900100, NET_OCR_INTERNAL = 991008 };
+
+template <typename T> inline T product(const std::vector<T> &vec) {
+  if (vec.empty()) {
+    return 0;
+  }
+  return std::accumulate(vec.begin(), vec.end(), 1, std::multiplies<T>());
+}

deploy/android_demo/app/src/main/cpp/native.cpp

+//
+// Created by fujiayi on 2020/7/5.
+//
+
+#include "native.h"
+#include "ocr_ppredictor.h"
+#include <algorithm>
+#include <paddle_api.h>
+#include <string>
+
+static paddle::lite_api::PowerMode str_to_cpu_mode(const std::string &cpu_mode);
+
+extern "C" JNIEXPORT jlong JNICALL
+Java_com_baidu_paddle_lite_demo_ocr_OCRPredictorNative_init(
+    JNIEnv *env, jobject thiz, jstring j_det_model_path,
+    jstring j_rec_model_path, jstring j_cls_model_path, jint j_thread_num,
+    jstring j_cpu_mode) {
+  std::string det_model_path = jstring_to_cpp_string(env, j_det_model_path);
+  std::string rec_model_path = jstring_to_cpp_string(env, j_rec_model_path);
+  std::string cls_model_path = jstring_to_cpp_string(env, j_cls_model_path);
+  int thread_num = j_thread_num;
+  std::string cpu_mode = jstring_to_cpp_string(env, j_cpu_mode);
+  ppredictor::OCR_Config conf;
+  conf.thread_num = thread_num;
+  conf.mode = str_to_cpu_mode(cpu_mode);
+  ppredictor::OCR_PPredictor *orc_predictor =
+      new ppredictor::OCR_PPredictor{conf};
+  orc_predictor->init_from_file(det_model_path, rec_model_path, cls_model_path);
+  return reinterpret_cast<jlong>(orc_predictor);
+}
+
+/**
+ * "LITE_POWER_HIGH" convert to paddle::lite_api::LITE_POWER_HIGH
+ * @param cpu_mode
+ * @return
+ */
+static paddle::lite_api::PowerMode
+str_to_cpu_mode(const std::string &cpu_mode) {
+  static std::map<std::string, paddle::lite_api::PowerMode> cpu_mode_map{
+      {"LITE_POWER_HIGH", paddle::lite_api::LITE_POWER_HIGH},
+      {"LITE_POWER_LOW", paddle::lite_api::LITE_POWER_HIGH},
+      {"LITE_POWER_FULL", paddle::lite_api::LITE_POWER_FULL},
+      {"LITE_POWER_NO_BIND", paddle::lite_api::LITE_POWER_NO_BIND},
+      {"LITE_POWER_RAND_HIGH", paddle::lite_api::LITE_POWER_RAND_HIGH},
+      {"LITE_POWER_RAND_LOW", paddle::lite_api::LITE_POWER_RAND_LOW}};
+  std::string upper_key;
+  std::transform(cpu_mode.cbegin(), cpu_mode.cend(), upper_key.begin(),
+                 ::toupper);
+  auto index = cpu_mode_map.find(upper_key);
+  if (index == cpu_mode_map.end()) {
+    LOGE("cpu_mode not found %s", upper_key.c_str());
+    return paddle::lite_api::LITE_POWER_HIGH;
+  } else {
+    return index->second;
+  }
+}
+
+extern "C" JNIEXPORT jfloatArray JNICALL
+Java_com_baidu_paddle_lite_demo_ocr_OCRPredictorNative_forward(
+    JNIEnv *env, jobject thiz, jlong java_pointer, jfloatArray buf,
+    jfloatArray ddims, jobject original_image) {
+  LOGI("begin to run native forward");
+  if (java_pointer == 0) {
+    LOGE("JAVA pointer is NULL");
+    return cpp_array_to_jfloatarray(env, nullptr, 0);
+  }
+  cv::Mat origin = bitmap_to_cv_mat(env, original_image);
+  if (origin.size == 0) {
+    LOGE("origin bitmap cannot convert to CV Mat");
+    return cpp_array_to_jfloatarray(env, nullptr, 0);
+  }
+  ppredictor::OCR_PPredictor *ppredictor =
+      (ppredictor::OCR_PPredictor *)java_pointer;
+  std::vector<float> dims_float_arr = jfloatarray_to_float_vector(env, ddims);
+  std::vector<int64_t> dims_arr;
+  dims_arr.resize(dims_float_arr.size());
+  std::copy(dims_float_arr.cbegin(), dims_float_arr.cend(), dims_arr.begin());
+
+  // 这里值有点大，就不调用jfloatarray_to_float_vector了
+  int64_t buf_len = (int64_t)env->GetArrayLength(buf);
+  jfloat *buf_data = env->GetFloatArrayElements(buf, JNI_FALSE);
+  float *data = (jfloat *)buf_data;
+  std::vector<ppredictor::OCRPredictResult> results =
+      ppredictor->infer_ocr(dims_arr, data, buf_len, NET_OCR, origin);
+  LOGI("infer_ocr finished with boxes %ld", results.size());
+  // 这里将std::vector<ppredictor::OCRPredictResult> 序列化成
+  // float数组，传输到java层再反序列化
+  std::vector<float> float_arr;
+  for (const ppredictor::OCRPredictResult &r : results) {
+    float_arr.push_back(r.points.size());
+    float_arr.push_back(r.word_index.size());
+    float_arr.push_back(r.score);
+    for (const std::vector<int> &point : r.points) {
+      float_arr.push_back(point.at(0));
+      float_arr.push_back(point.at(1));
+    }
+    for (int index : r.word_index) {
+      float_arr.push_back(index);
+    }
+  }
+  return cpp_array_to_jfloatarray(env, float_arr.data(), float_arr.size());
+}
+
+extern "C" JNIEXPORT void JNICALL
+Java_com_baidu_paddle_lite_demo_ocr_OCRPredictorNative_release(
+    JNIEnv *env, jobject thiz, jlong java_pointer) {
+  if (java_pointer == 0) {
+    LOGE("JAVA pointer is NULL");
+    return;
+  }
+  ppredictor::OCR_PPredictor *ppredictor =
+      (ppredictor::OCR_PPredictor *)java_pointer;
+  delete ppredictor;
+}
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/native.h

+//
+// Created by fujiayi on 2020/7/5.
+//
+
+#pragma once
+
+#include "common.h"
+#include <android/bitmap.h>
+#include <jni.h>
+#include <opencv2/opencv.hpp>
+#include <string>
+#include <vector>
+
+inline std::string jstring_to_cpp_string(JNIEnv *env, jstring jstr) {
+  // In java, a unicode char will be encoded using 2 bytes (utf16).
+  // so jstring will contain characters utf16. std::string in c++ is
+  // essentially a string of bytes, not characters, so if we want to
+  // pass jstring from JNI to c++, we have convert utf16 to bytes.
+  if (!jstr) {
+    return "";
+  }
+  const jclass stringClass = env->GetObjectClass(jstr);
+  const jmethodID getBytes =
+      env->GetMethodID(stringClass, "getBytes", "(Ljava/lang/String;)[B");
+  const jbyteArray stringJbytes = (jbyteArray)env->CallObjectMethod(
+      jstr, getBytes, env->NewStringUTF("UTF-8"));
+
+  size_t length = (size_t)env->GetArrayLength(stringJbytes);
+  jbyte *pBytes = env->GetByteArrayElements(stringJbytes, NULL);
+
+  std::string ret = std::string(reinterpret_cast<char *>(pBytes), length);
+  env->ReleaseByteArrayElements(stringJbytes, pBytes, JNI_ABORT);
+
+  env->DeleteLocalRef(stringJbytes);
+  env->DeleteLocalRef(stringClass);
+  return ret;
+}
+
+inline jstring cpp_string_to_jstring(JNIEnv *env, std::string str) {
+  auto *data = str.c_str();
+  jclass strClass = env->FindClass("java/lang/String");
+  jmethodID strClassInitMethodID =
+      env->GetMethodID(strClass, "<init>", "([BLjava/lang/String;)V");
+
+  jbyteArray bytes = env->NewByteArray(strlen(data));
+  env->SetByteArrayRegion(bytes, 0, strlen(data),
+                          reinterpret_cast<const jbyte *>(data));
+
+  jstring encoding = env->NewStringUTF("UTF-8");
+  jstring res = (jstring)(
+      env->NewObject(strClass, strClassInitMethodID, bytes, encoding));
+
+  env->DeleteLocalRef(strClass);
+  env->DeleteLocalRef(encoding);
+  env->DeleteLocalRef(bytes);
+
+  return res;
+}
+
+inline jfloatArray cpp_array_to_jfloatarray(JNIEnv *env, const float *buf,
+                                            int64_t len) {
+  if (len == 0) {
+    return env->NewFloatArray(0);
+  }
+  jfloatArray result = env->NewFloatArray(len);
+  env->SetFloatArrayRegion(result, 0, len, buf);
+  return result;
+}
+
+inline jintArray cpp_array_to_jintarray(JNIEnv *env, const int *buf,
+                                        int64_t len) {
+  jintArray result = env->NewIntArray(len);
+  env->SetIntArrayRegion(result, 0, len, buf);
+  return result;
+}
+
+inline jbyteArray cpp_array_to_jbytearray(JNIEnv *env, const int8_t *buf,
+                                          int64_t len) {
+  jbyteArray result = env->NewByteArray(len);
+  env->SetByteArrayRegion(result, 0, len, buf);
+  return result;
+}
+
+inline jlongArray int64_vector_to_jlongarray(JNIEnv *env,
+                                             const std::vector<int64_t> &vec) {
+  jlongArray result = env->NewLongArray(vec.size());
+  jlong *buf = new jlong[vec.size()];
+  for (size_t i = 0; i < vec.size(); ++i) {
+    buf[i] = (jlong)vec[i];
+  }
+  env->SetLongArrayRegion(result, 0, vec.size(), buf);
+  delete[] buf;
+  return result;
+}
+
+inline std::vector<int64_t> jlongarray_to_int64_vector(JNIEnv *env,
+                                                       jlongArray data) {
+  int data_size = env->GetArrayLength(data);
+  jlong *data_ptr = env->GetLongArrayElements(data, nullptr);
+  std::vector<int64_t> data_vec(data_ptr, data_ptr + data_size);
+  env->ReleaseLongArrayElements(data, data_ptr, 0);
+  return data_vec;
+}
+
+inline std::vector<float> jfloatarray_to_float_vector(JNIEnv *env,
+                                                      jfloatArray data) {
+  int data_size = env->GetArrayLength(data);
+  jfloat *data_ptr = env->GetFloatArrayElements(data, nullptr);
+  std::vector<float> data_vec(data_ptr, data_ptr + data_size);
+  env->ReleaseFloatArrayElements(data, data_ptr, 0);
+  return data_vec;
+}
+
+inline cv::Mat bitmap_to_cv_mat(JNIEnv *env, jobject bitmap) {
+  AndroidBitmapInfo info;
+  int result = AndroidBitmap_getInfo(env, bitmap, &info);
+  if (result != ANDROID_BITMAP_RESULT_SUCCESS) {
+    LOGE("AndroidBitmap_getInfo failed, result: %d", result);
+    return cv::Mat{};
+  }
+  if (info.format != ANDROID_BITMAP_FORMAT_RGBA_8888) {
+    LOGE("Bitmap format is not RGBA_8888 !");
+    return cv::Mat{};
+  }
+  unsigned char *srcData = NULL;
+  AndroidBitmap_lockPixels(env, bitmap, (void **)&srcData);
+  cv::Mat mat = cv::Mat::zeros(info.height, info.width, CV_8UC4);
+  memcpy(mat.data, srcData, info.height * info.width * 4);
+  AndroidBitmap_unlockPixels(env, bitmap);
+  cv::cvtColor(mat, mat, cv::COLOR_RGBA2BGR);
+  /**
+  if (!cv::imwrite("/sdcard/1/copy.jpg", mat)){
+      LOGE("Write image failed " );
+  }
+   */
+  return mat;
+}

deploy/android_demo/app/src/main/cpp/ocr_clipper.hpp

+/*******************************************************************************
+*                                                                              *
+* Author    :  Angus Johnson                                                   *
+* Version   :  6.4.2                                                           *
+* Date      :  27 February 2017                                                *
+* Website   :  http://www.angusj.com                                           *
+* Copyright :  Angus Johnson 2010-2017                                         *
+*                                                                              *
+* License:                                                                     *
+* Use, modification & distribution is subject to Boost Software License Ver 1. *
+* http://www.boost.org/LICENSE_1_0.txt                                         *
+*                                                                              *
+* Attributions:                                                                *
+* The code in this library is an extension of Bala Vatti's clipping algorithm: *
+* "A generic solution to polygon clipping"                                     *
+* Communications of the ACM, Vol 35, Issue 7 (July 1992) pp 56-63.             *
+* http://portal.acm.org/citation.cfm?id=129906                                 *
+*                                                                              *
+* Computer graphics and geometric modeling: implementation and algorithms      *
+* By Max K. Agoston                                                            *
+* Springer; 1 edition (January 4, 2005)                                        *
+* http://books.google.com/books?q=vatti+clipping+agoston                       *
+*                                                                              *
+* See also:                                                                    *
+* "Polygon Offsetting by Computing Winding Numbers"                            *
+* Paper no. DETC2005-85513 pp. 565-575                                         *
+* ASME 2005 International Design Engineering Technical Conferences             *
+* and Computers and Information in Engineering Conference (IDETC/CIE2005)      *
+* September 24-28, 2005 , Long Beach, California, USA                          *
+* http://www.me.berkeley.edu/~mcmains/pubs/DAC05OffsetPolygon.pdf              *
+*                                                                              *
+*******************************************************************************/
+
+#ifndef clipper_hpp
+#define clipper_hpp
+
+#define CLIPPER_VERSION "6.4.2"
+
+// use_int32: When enabled 32bit ints are used instead of 64bit ints. This
+// improve performance but coordinate values are limited to the range +/- 46340
+//#define use_int32
+
+// use_xyz: adds a Z member to IntPoint. Adds a minor cost to perfomance.
+//#define use_xyz
+
+// use_lines: Enables line clipping. Adds a very minor cost to performance.
+#define use_lines
+
+// use_deprecated: Enables temporary support for the obsolete functions
+//#define use_deprecated
+
+#include <cstdlib>
+#include <cstring>
+#include <functional>
+#include <list>
+#include <ostream>
+#include <queue>
+#include <set>
+#include <stdexcept>
+#include <vector>
+
+namespace ClipperLib {
+
+enum ClipType { ctIntersection, ctUnion, ctDifference, ctXor };
+enum PolyType { ptSubject, ptClip };
+// By far the most widely used winding rules for polygon filling are
+// EvenOdd & NonZero (GDI, GDI+, XLib, OpenGL, Cairo, AGG, Quartz, SVG, Gr32)
+// Others rules include Positive, Negative and ABS_GTR_EQ_TWO (only in OpenGL)
+// see http://glprogramming.com/red/chapter11.html
+enum PolyFillType { pftEvenOdd, pftNonZero, pftPositive, pftNegative };
+
+#ifdef use_int32
+typedef int cInt;
+static cInt const loRange = 0x7FFF;
+static cInt const hiRange = 0x7FFF;
+#else
+typedef signed long long cInt;
+static cInt const loRange = 0x3FFFFFFF;
+static cInt const hiRange = 0x3FFFFFFFFFFFFFFFLL;
+typedef signed long long long64; // used by Int128 class
+typedef unsigned long long ulong64;
+
+#endif
+
+struct IntPoint {
+  cInt X;
+  cInt Y;
+#ifdef use_xyz
+  cInt Z;
+  IntPoint(cInt x = 0, cInt y = 0, cInt z = 0) : X(x), Y(y), Z(z){};
+#else
+
+  IntPoint(cInt x = 0, cInt y = 0) : X(x), Y(y){};
+#endif
+
+  friend inline bool operator==(const IntPoint &a, const IntPoint &b) {
+    return a.X == b.X && a.Y == b.Y;
+  }
+
+  friend inline bool operator!=(const IntPoint &a, const IntPoint &b) {
+    return a.X != b.X || a.Y != b.Y;
+  }
+};
+//------------------------------------------------------------------------------
+
+typedef std::vector<IntPoint> Path;
+typedef std::vector<Path> Paths;
+
+inline Path &operator<<(Path &poly, const IntPoint &p) {
+  poly.push_back(p);
+  return poly;
+}
+
+inline Paths &operator<<(Paths &polys, const Path &p) {
+  polys.push_back(p);
+  return polys;
+}
+
+std::ostream &operator<<(std::ostream &s, const IntPoint &p);
+
+std::ostream &operator<<(std::ostream &s, const Path &p);
+
+std::ostream &operator<<(std::ostream &s, const Paths &p);
+
+struct DoublePoint {
+  double X;
+  double Y;
+
+  DoublePoint(double x = 0, double y = 0) : X(x), Y(y) {}
+
+  DoublePoint(IntPoint ip) : X((double)ip.X), Y((double)ip.Y) {}
+};
+//------------------------------------------------------------------------------
+
+#ifdef use_xyz
+typedef void (*ZFillCallback)(IntPoint &e1bot, IntPoint &e1top, IntPoint &e2bot,
+                              IntPoint &e2top, IntPoint &pt);
+#endif
+
+enum InitOptions {
+  ioReverseSolution = 1,
+  ioStrictlySimple = 2,
+  ioPreserveCollinear = 4
+};
+enum JoinType { jtSquare, jtRound, jtMiter };
+enum EndType {
+  etClosedPolygon,
+  etClosedLine,
+  etOpenButt,
+  etOpenSquare,
+  etOpenRound
+};
+
+class PolyNode;
+
+typedef std::vector<PolyNode *> PolyNodes;
+
+class PolyNode {
+public:
+  PolyNode();
+
+  virtual ~PolyNode(){};
+  Path Contour;
+  PolyNodes Childs;
+  PolyNode *Parent;
+
+  PolyNode *GetNext() const;
+
+  bool IsHole() const;
+
+  bool IsOpen() const;
+
+  int ChildCount() const;
+
+private:
+  // PolyNode& operator =(PolyNode& other);
+  unsigned Index; // node index in Parent.Childs
+  bool m_IsOpen;
+  JoinType m_jointype;
+  EndType m_endtype;
+
+  PolyNode *GetNextSiblingUp() const;
+
+  void AddChild(PolyNode &child);
+
+  friend class Clipper; // to access Index
+  friend class ClipperOffset;
+};
+
+class PolyTree : public PolyNode {
+public:
+  ~PolyTree() { Clear(); };
+
+  PolyNode *GetFirst() const;
+
+  void Clear();
+
+  int Total() const;
+
+private:
+  // PolyTree& operator =(PolyTree& other);
+  PolyNodes AllNodes;
+
+  friend class Clipper; // to access AllNodes
+};
+
+bool Orientation(const Path &poly);
+
+double Area(const Path &poly);
+
+int PointInPolygon(const IntPoint &pt, const Path &path);
+
+void SimplifyPolygon(const Path &in_poly, Paths &out_polys,
+                     PolyFillType fillType = pftEvenOdd);
+
+void SimplifyPolygons(const Paths &in_polys, Paths &out_polys,
+                      PolyFillType fillType = pftEvenOdd);
+
+void SimplifyPolygons(Paths &polys, PolyFillType fillType = pftEvenOdd);
+
+void CleanPolygon(const Path &in_poly, Path &out_poly, double distance = 1.415);
+
+void CleanPolygon(Path &poly, double distance = 1.415);
+
+void CleanPolygons(const Paths &in_polys, Paths &out_polys,
+                   double distance = 1.415);
+
+void CleanPolygons(Paths &polys, double distance = 1.415);
+
+void MinkowskiSum(const Path &pattern, const Path &path, Paths &solution,
+                  bool pathIsClosed);
+
+void MinkowskiSum(const Path &pattern, const Paths &paths, Paths &solution,
+                  bool pathIsClosed);
+
+void MinkowskiDiff(const Path &poly1, const Path &poly2, Paths &solution);
+
+void PolyTreeToPaths(const PolyTree &polytree, Paths &paths);
+
+void ClosedPathsFromPolyTree(const PolyTree &polytree, Paths &paths);
+
+void OpenPathsFromPolyTree(PolyTree &polytree, Paths &paths);
+
+void ReversePath(Path &p);
+
+void ReversePaths(Paths &p);
+
+struct IntRect {
+  cInt left;
+  cInt top;
+  cInt right;
+  cInt bottom;
+};
+
+// enums that are used internally ...
+enum EdgeSide { esLeft = 1, esRight = 2 };
+
+// forward declarations (for stuff used internally) ...
+struct TEdge;
+struct IntersectNode;
+struct LocalMinimum;
+struct OutPt;
+struct OutRec;
+struct Join;
+
+typedef std::vector<OutRec *> PolyOutList;
+typedef std::vector<TEdge *> EdgeList;
+typedef std::vector<Join *> JoinList;
+typedef std::vector<IntersectNode *> IntersectList;
+
+//------------------------------------------------------------------------------
+
+// ClipperBase is the ancestor to the Clipper class. It should not be
+// instantiated directly. This class simply abstracts the conversion of sets of
+// polygon coordinates into edge objects that are stored in a LocalMinima list.
+class ClipperBase {
+public:
+  ClipperBase();
+
+  virtual ~ClipperBase();
+
+  virtual bool AddPath(const Path &pg, PolyType PolyTyp, bool Closed);
+
+  bool AddPaths(const Paths &ppg, PolyType PolyTyp, bool Closed);
+
+  virtual void Clear();
+
+  IntRect GetBounds();
+
+  bool PreserveCollinear() { return m_PreserveCollinear; };
+
+  void PreserveCollinear(bool value) { m_PreserveCollinear = value; };
+
+protected:
+  void DisposeLocalMinimaList();
+
+  TEdge *AddBoundsToLML(TEdge *e, bool IsClosed);
+
+  virtual void Reset();
+
+  TEdge *ProcessBound(TEdge *E, bool IsClockwise);
+
+  void InsertScanbeam(const cInt Y);
+
+  bool PopScanbeam(cInt &Y);
+
+  bool LocalMinimaPending();
+
+  bool PopLocalMinima(cInt Y, const LocalMinimum *&locMin);
+
+  OutRec *CreateOutRec();
+
+  void DisposeAllOutRecs();
+
+  void DisposeOutRec(PolyOutList::size_type index);
+
+  void SwapPositionsInAEL(TEdge *edge1, TEdge *edge2);
+
+  void DeleteFromAEL(TEdge *e);
+
+  void UpdateEdgeIntoAEL(TEdge *&e);
+
+  typedef std::vector<LocalMinimum> MinimaList;
+  MinimaList::iterator m_CurrentLM;
+  MinimaList m_MinimaList;
+
+  bool m_UseFullRange;
+  EdgeList m_edges;
+  bool m_PreserveCollinear;
+  bool m_HasOpenPaths;
+  PolyOutList m_PolyOuts;
+  TEdge *m_ActiveEdges;
+
+  typedef std::priority_queue<cInt> ScanbeamList;
+  ScanbeamList m_Scanbeam;
+};
+//------------------------------------------------------------------------------
+
+class Clipper : public virtual ClipperBase {
+public:
+  Clipper(int initOptions = 0);
+
+  bool Execute(ClipType clipType, Paths &solution,
+               PolyFillType fillType = pftEvenOdd);
+
+  bool Execute(ClipType clipType, Paths &solution, PolyFillType subjFillType,
+               PolyFillType clipFillType);
+
+  bool Execute(ClipType clipType, PolyTree &polytree,
+               PolyFillType fillType = pftEvenOdd);
+
+  bool Execute(ClipType clipType, PolyTree &polytree, PolyFillType subjFillType,
+               PolyFillType clipFillType);
+
+  bool ReverseSolution() { return m_ReverseOutput; };
+
+  void ReverseSolution(bool value) { m_ReverseOutput = value; };
+
+  bool StrictlySimple() { return m_StrictSimple; };
+
+  void StrictlySimple(bool value) { m_StrictSimple = value; };
+// set the callback function for z value filling on intersections (otherwise Z
+// is 0)
+#ifdef use_xyz
+  void ZFillFunction(ZFillCallback zFillFunc);
+#endif
+protected:
+  virtual bool ExecuteInternal();
+
+private:
+  JoinList m_Joins;
+  JoinList m_GhostJoins;
+  IntersectList m_IntersectList;
+  ClipType m_ClipType;
+  typedef std::list<cInt> MaximaList;
+  MaximaList m_Maxima;
+  TEdge *m_SortedEdges;
+  bool m_ExecuteLocked;
+  PolyFillType m_ClipFillType;
+  PolyFillType m_SubjFillType;
+  bool m_ReverseOutput;
+  bool m_UsingPolyTree;
+  bool m_StrictSimple;
+#ifdef use_xyz
+  ZFillCallback m_ZFill; // custom callback
+#endif
+
+  void SetWindingCount(TEdge &edge);
+
+  bool IsEvenOddFillType(const TEdge &edge) const;
+
+  bool IsEvenOddAltFillType(const TEdge &edge) const;
+
+  void InsertLocalMinimaIntoAEL(const cInt botY);
+
+  void InsertEdgeIntoAEL(TEdge *edge, TEdge *startEdge);
+
+  void AddEdgeToSEL(TEdge *edge);
+
+  bool PopEdgeFromSEL(TEdge *&edge);
+
+  void CopyAELToSEL();
+
+  void DeleteFromSEL(TEdge *e);
+
+  void SwapPositionsInSEL(TEdge *edge1, TEdge *edge2);
+
+  bool IsContributing(const TEdge &edge) const;
+
+  bool IsTopHorz(const cInt XPos);
+
+  void DoMaxima(TEdge *e);
+
+  void ProcessHorizontals();
+
+  void ProcessHorizontal(TEdge *horzEdge);
+
+  void AddLocalMaxPoly(TEdge *e1, TEdge *e2, const IntPoint &pt);
+
+  OutPt *AddLocalMinPoly(TEdge *e1, TEdge *e2, const IntPoint &pt);
+
+  OutRec *GetOutRec(int idx);
+
+  void AppendPolygon(TEdge *e1, TEdge *e2);
+
+  void IntersectEdges(TEdge *e1, TEdge *e2, IntPoint &pt);
+
+  OutPt *AddOutPt(TEdge *e, const IntPoint &pt);
+
+  OutPt *GetLastOutPt(TEdge *e);
+
+  bool ProcessIntersections(const cInt topY);
+
+  void BuildIntersectList(const cInt topY);
+
+  void ProcessIntersectList();
+
+  void ProcessEdgesAtTopOfScanbeam(const cInt topY);
+
+  void BuildResult(Paths &polys);
+
+  void BuildResult2(PolyTree &polytree);
+
+  void SetHoleState(TEdge *e, OutRec *outrec);
+
+  void DisposeIntersectNodes();
+
+  bool FixupIntersectionOrder();
+
+  void FixupOutPolygon(OutRec &outrec);
+
+  void FixupOutPolyline(OutRec &outrec);
+
+  bool IsHole(TEdge *e);
+
+  bool FindOwnerFromSplitRecs(OutRec &outRec, OutRec *&currOrfl);
+
+  void FixHoleLinkage(OutRec &outrec);
+
+  void AddJoin(OutPt *op1, OutPt *op2, const IntPoint offPt);
+
+  void ClearJoins();
+
+  void ClearGhostJoins();
+
+  void AddGhostJoin(OutPt *op, const IntPoint offPt);
+
+  bool JoinPoints(Join *j, OutRec *outRec1, OutRec *outRec2);
+
+  void JoinCommonEdges();
+
+  void DoSimplePolygons();
+
+  void FixupFirstLefts1(OutRec *OldOutRec, OutRec *NewOutRec);
+
+  void FixupFirstLefts2(OutRec *InnerOutRec, OutRec *OuterOutRec);
+
+  void FixupFirstLefts3(OutRec *OldOutRec, OutRec *NewOutRec);
+
+#ifdef use_xyz
+  void SetZ(IntPoint &pt, TEdge &e1, TEdge &e2);
+#endif
+};
+//------------------------------------------------------------------------------
+
+class ClipperOffset {
+public:
+  ClipperOffset(double miterLimit = 2.0, double roundPrecision = 0.25);
+
+  ~ClipperOffset();
+
+  void AddPath(const Path &path, JoinType joinType, EndType endType);
+
+  void AddPaths(const Paths &paths, JoinType joinType, EndType endType);
+
+  void Execute(Paths &solution, double delta);
+
+  void Execute(PolyTree &solution, double delta);
+
+  void Clear();
+
+  double MiterLimit;
+  double ArcTolerance;
+
+private:
+  Paths m_destPolys;
+  Path m_srcPoly;
+  Path m_destPoly;
+  std::vector<DoublePoint> m_normals;
+  double m_delta, m_sinA, m_sin, m_cos;
+  double m_miterLim, m_StepsPerRad;
+  IntPoint m_lowest;
+  PolyNode m_polyNodes;
+
+  void FixOrientations();
+
+  void DoOffset(double delta);
+
+  void OffsetPoint(int j, int &k, JoinType jointype);
+
+  void DoSquare(int j, int k);
+
+  void DoMiter(int j, int k, double r);
+
+  void DoRound(int j, int k);
+};
+//------------------------------------------------------------------------------
+
+class clipperException : public std::exception {
+public:
+  clipperException(const char *description) : m_descr(description) {}
+
+  virtual ~clipperException() throw() {}
+
+  virtual const char *what() const throw() { return m_descr.c_str(); }
+
+private:
+  std::string m_descr;
+};
+//------------------------------------------------------------------------------
+
+} // ClipperLib namespace
+
+#endif // clipper_hpp

deploy/android_demo/app/src/main/cpp/ocr_cls_process.cpp

+// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "ocr_cls_process.h"
+#include <cmath>
+#include <cstring>
+#include <fstream>
+#include <iostream>
+#include <iostream>
+#include <vector>
+
+const std::vector<int> CLS_IMAGE_SHAPE = {3, 48, 192};
+
+cv::Mat cls_resize_img(const cv::Mat &img) {
+  int imgC = CLS_IMAGE_SHAPE[0];
+  int imgW = CLS_IMAGE_SHAPE[2];
+  int imgH = CLS_IMAGE_SHAPE[1];
+
+  float ratio = float(img.cols) / float(img.rows);
+  int resize_w = 0;
+  if (ceilf(imgH * ratio) > imgW)
+    resize_w = imgW;
+  else
+    resize_w = int(ceilf(imgH * ratio));
+
+  cv::Mat resize_img;
+  cv::resize(img, resize_img, cv::Size(resize_w, imgH), 0.f, 0.f,
+             cv::INTER_CUBIC);
+
+  if (resize_w < imgW) {
+    cv::copyMakeBorder(resize_img, resize_img, 0, 0, 0, int(imgW - resize_w),
+                       cv::BORDER_CONSTANT, {0, 0, 0});
+  }
+  return resize_img;
+}
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/ocr_cls_process.h

+// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "common.h"
+#include <opencv2/opencv.hpp>
+#include <vector>
+
+extern const std::vector<int> CLS_IMAGE_SHAPE;
+
+cv::Mat cls_resize_img(const cv::Mat &img);
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/ocr_crnn_process.cpp

+// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "ocr_crnn_process.h"
+#include <cmath>
+#include <cstring>
+#include <fstream>
+#include <iostream>
+#include <iostream>
+#include <vector>
+
+const std::string CHARACTER_TYPE = "ch";
+const int MAX_DICT_LENGTH = 6624;
+const std::vector<int> REC_IMAGE_SHAPE = {3, 32, 320};
+
+static cv::Mat crnn_resize_norm_img(cv::Mat img, float wh_ratio) {
+  int imgC = REC_IMAGE_SHAPE[0];
+  int imgW = REC_IMAGE_SHAPE[2];
+  int imgH = REC_IMAGE_SHAPE[1];
+
+  if (CHARACTER_TYPE == "ch")
+    imgW = int(32 * wh_ratio);
+
+  float ratio = float(img.cols) / float(img.rows);
+  int resize_w = 0;
+  if (ceilf(imgH * ratio) > imgW)
+    resize_w = imgW;
+  else
+    resize_w = int(ceilf(imgH * ratio));
+  cv::Mat resize_img;
+  cv::resize(img, resize_img, cv::Size(resize_w, imgH), 0.f, 0.f,
+             cv::INTER_CUBIC);
+
+  resize_img.convertTo(resize_img, CV_32FC3, 1 / 255.f);
+
+  for (int h = 0; h < resize_img.rows; h++) {
+    for (int w = 0; w < resize_img.cols; w++) {
+      resize_img.at<cv::Vec3f>(h, w)[0] =
+          (resize_img.at<cv::Vec3f>(h, w)[0] - 0.5) * 2;
+      resize_img.at<cv::Vec3f>(h, w)[1] =
+          (resize_img.at<cv::Vec3f>(h, w)[1] - 0.5) * 2;
+      resize_img.at<cv::Vec3f>(h, w)[2] =
+          (resize_img.at<cv::Vec3f>(h, w)[2] - 0.5) * 2;
+    }
+  }
+
+  cv::Mat dist;
+  cv::copyMakeBorder(resize_img, dist, 0, 0, 0, int(imgW - resize_w),
+                     cv::BORDER_CONSTANT, {0, 0, 0});
+
+  return dist;
+}
+
+cv::Mat crnn_resize_img(const cv::Mat &img, float wh_ratio) {
+  int imgC = REC_IMAGE_SHAPE[0];
+  int imgW = REC_IMAGE_SHAPE[2];
+  int imgH = REC_IMAGE_SHAPE[1];
+
+  if (CHARACTER_TYPE == "ch") {
+    imgW = int(32 * wh_ratio);
+  }
+
+  float ratio = float(img.cols) / float(img.rows);
+  int resize_w = 0;
+  if (ceilf(imgH * ratio) > imgW)
+    resize_w = imgW;
+  else
+    resize_w = int(ceilf(imgH * ratio));
+  cv::Mat resize_img;
+  cv::resize(img, resize_img, cv::Size(resize_w, imgH));
+  return resize_img;
+}
+
+cv::Mat get_rotate_crop_image(const cv::Mat &srcimage,
+                              const std::vector<std::vector<int>> &box) {
+
+  std::vector<std::vector<int>> points = box;
+
+  int x_collect[4] = {box[0][0], box[1][0], box[2][0], box[3][0]};
+  int y_collect[4] = {box[0][1], box[1][1], box[2][1], box[3][1]};
+  int left = int(*std::min_element(x_collect, x_collect + 4));
+  int right = int(*std::max_element(x_collect, x_collect + 4));
+  int top = int(*std::min_element(y_collect, y_collect + 4));
+  int bottom = int(*std::max_element(y_collect, y_collect + 4));
+
+  cv::Mat img_crop;
+  srcimage(cv::Rect(left, top, right - left, bottom - top)).copyTo(img_crop);
+
+  for (int i = 0; i < points.size(); i++) {
+    points[i][0] -= left;
+    points[i][1] -= top;
+  }
+
+  int img_crop_width = int(sqrt(pow(points[0][0] - points[1][0], 2) +
+                                pow(points[0][1] - points[1][1], 2)));
+  int img_crop_height = int(sqrt(pow(points[0][0] - points[3][0], 2) +
+                                 pow(points[0][1] - points[3][1], 2)));
+
+  cv::Point2f pts_std[4];
+  pts_std[0] = cv::Point2f(0., 0.);
+  pts_std[1] = cv::Point2f(img_crop_width, 0.);
+  pts_std[2] = cv::Point2f(img_crop_width, img_crop_height);
+  pts_std[3] = cv::Point2f(0.f, img_crop_height);
+
+  cv::Point2f pointsf[4];
+  pointsf[0] = cv::Point2f(points[0][0], points[0][1]);
+  pointsf[1] = cv::Point2f(points[1][0], points[1][1]);
+  pointsf[2] = cv::Point2f(points[2][0], points[2][1]);
+  pointsf[3] = cv::Point2f(points[3][0], points[3][1]);
+
+  cv::Mat M = cv::getPerspectiveTransform(pointsf, pts_std);
+
+  cv::Mat dst_img;
+  cv::warpPerspective(img_crop, dst_img, M,
+                      cv::Size(img_crop_width, img_crop_height),
+                      cv::BORDER_REPLICATE);
+
+  if (float(dst_img.rows) >= float(dst_img.cols) * 1.5) {
+    /*
+    cv::Mat srcCopy = cv::Mat(dst_img.rows, dst_img.cols, dst_img.depth());
+    cv::transpose(dst_img, srcCopy);
+    cv::flip(srcCopy, srcCopy, 0);
+    return srcCopy;
+    */
+    cv::transpose(dst_img, dst_img);
+    cv::flip(dst_img, dst_img, 0);
+    return dst_img;
+  } else {
+    return dst_img;
+  }
+}

deploy/android_demo/app/src/main/cpp/ocr_crnn_process.h

+//
+// Created by fujiayi on 2020/7/3.
+//
+#pragma once
+
+#include "common.h"
+#include <opencv2/opencv.hpp>
+#include <vector>
+
+extern const std::vector<int> REC_IMAGE_SHAPE;
+
+cv::Mat get_rotate_crop_image(const cv::Mat &srcimage,
+                              const std::vector<std::vector<int>> &box);
+
+cv::Mat crnn_resize_img(const cv::Mat &img, float wh_ratio);
+
+template <class ForwardIterator>
+inline size_t argmax(ForwardIterator first, ForwardIterator last) {
+  return std::distance(first, std::max_element(first, last));
+}
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/ocr_db_post_process.cpp

+// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "ocr_clipper.hpp"
+#include "opencv2/core.hpp"
+#include "opencv2/imgcodecs.hpp"
+#include "opencv2/imgproc.hpp"
+#include <iostream>
+#include <math.h>
+#include <vector>
+
+static void getcontourarea(float **box, float unclip_ratio, float &distance) {
+  int pts_num = 4;
+  float area = 0.0f;
+  float dist = 0.0f;
+  for (int i = 0; i < pts_num; i++) {
+    area += box[i][0] * box[(i + 1) % pts_num][1] -
+            box[i][1] * box[(i + 1) % pts_num][0];
+    dist += sqrtf((box[i][0] - box[(i + 1) % pts_num][0]) *
+                      (box[i][0] - box[(i + 1) % pts_num][0]) +
+                  (box[i][1] - box[(i + 1) % pts_num][1]) *
+                      (box[i][1] - box[(i + 1) % pts_num][1]));
+  }
+  area = fabs(float(area / 2.0));
+
+  distance = area * unclip_ratio / dist;
+}
+
+static cv::RotatedRect unclip(float **box) {
+  float unclip_ratio = 2.0;
+  float distance = 1.0;
+
+  getcontourarea(box, unclip_ratio, distance);
+
+  ClipperLib::ClipperOffset offset;
+  ClipperLib::Path p;
+  p << ClipperLib::IntPoint(int(box[0][0]), int(box[0][1]))
+    << ClipperLib::IntPoint(int(box[1][0]), int(box[1][1]))
+    << ClipperLib::IntPoint(int(box[2][0]), int(box[2][1]))
+    << ClipperLib::IntPoint(int(box[3][0]), int(box[3][1]));
+  offset.AddPath(p, ClipperLib::jtRound, ClipperLib::etClosedPolygon);
+
+  ClipperLib::Paths soln;
+  offset.Execute(soln, distance);
+  std::vector<cv::Point2f> points;
+
+  for (int j = 0; j < soln.size(); j++) {
+    for (int i = 0; i < soln[soln.size() - 1].size(); i++) {
+      points.emplace_back(soln[j][i].X, soln[j][i].Y);
+    }
+  }
+  cv::RotatedRect res = cv::minAreaRect(points);
+
+  return res;
+}
+
+static float **Mat2Vec(cv::Mat mat) {
+  auto **array = new float *[mat.rows];
+  for (int i = 0; i < mat.rows; ++i) {
+    array[i] = new float[mat.cols];
+  }
+  for (int i = 0; i < mat.rows; ++i) {
+    for (int j = 0; j < mat.cols; ++j) {
+      array[i][j] = mat.at<float>(i, j);
+    }
+  }
+
+  return array;
+}
+
+static void quickSort(float **s, int l, int r) {
+  if (l < r) {
+    int i = l, j = r;
+    float x = s[l][0];
+    float *xp = s[l];
+    while (i < j) {
+      while (i < j && s[j][0] >= x) {
+        j--;
+      }
+      if (i < j) {
+        std::swap(s[i++], s[j]);
+      }
+      while (i < j && s[i][0] < x) {
+        i++;
+      }
+      if (i < j) {
+        std::swap(s[j--], s[i]);
+      }
+    }
+    s[i] = xp;
+    quickSort(s, l, i - 1);
+    quickSort(s, i + 1, r);
+  }
+}
+
+static void quickSort_vector(std::vector<std::vector<int>> &box, int l, int r,
+                             int axis) {
+  if (l < r) {
+    int i = l, j = r;
+    int x = box[l][axis];
+    std::vector<int> xp(box[l]);
+    while (i < j) {
+      while (i < j && box[j][axis] >= x) {
+        j--;
+      }
+      if (i < j) {
+        std::swap(box[i++], box[j]);
+      }
+      while (i < j && box[i][axis] < x) {
+        i++;
+      }
+      if (i < j) {
+        std::swap(box[j--], box[i]);
+      }
+    }
+    box[i] = xp;
+    quickSort_vector(box, l, i - 1, axis);
+    quickSort_vector(box, i + 1, r, axis);
+  }
+}
+
+static std::vector<std::vector<int>>
+order_points_clockwise(std::vector<std::vector<int>> pts) {
+  std::vector<std::vector<int>> box = pts;
+  quickSort_vector(box, 0, int(box.size() - 1), 0);
+  std::vector<std::vector<int>> leftmost = {box[0], box[1]};
+  std::vector<std::vector<int>> rightmost = {box[2], box[3]};
+
+  if (leftmost[0][1] > leftmost[1][1]) {
+    std::swap(leftmost[0], leftmost[1]);
+  }
+
+  if (rightmost[0][1] > rightmost[1][1]) {
+    std::swap(rightmost[0], rightmost[1]);
+  }
+
+  std::vector<std::vector<int>> rect = {leftmost[0], rightmost[0], rightmost[1],
+                                        leftmost[1]};
+  return rect;
+}
+
+static float **get_mini_boxes(cv::RotatedRect box, float &ssid) {
+  ssid = box.size.width >= box.size.height ? box.size.height : box.size.width;
+
+  cv::Mat points;
+  cv::boxPoints(box, points);
+  // sorted box points
+  auto array = Mat2Vec(points);
+  quickSort(array, 0, 3);
+
+  float *idx1 = array[0], *idx2 = array[1], *idx3 = array[2], *idx4 = array[3];
+  if (array[3][1] <= array[2][1]) {
+    idx2 = array[3];
+    idx3 = array[2];
+  } else {
+    idx2 = array[2];
+    idx3 = array[3];
+  }
+  if (array[1][1] <= array[0][1]) {
+    idx1 = array[1];
+    idx4 = array[0];
+  } else {
+    idx1 = array[0];
+    idx4 = array[1];
+  }
+
+  array[0] = idx1;
+  array[1] = idx2;
+  array[2] = idx3;
+  array[3] = idx4;
+
+  return array;
+}
+
+template <class T> T clamp(T x, T min, T max) {
+  if (x > max) {
+    return max;
+  }
+  if (x < min) {
+    return min;
+  }
+  return x;
+}
+
+static float clampf(float x, float min, float max) {
+  if (x > max)
+    return max;
+  if (x < min)
+    return min;
+  return x;
+}
+
+float box_score_fast(float **box_array, cv::Mat pred) {
+  auto array = box_array;
+  int width = pred.cols;
+  int height = pred.rows;
+
+  float box_x[4] = {array[0][0], array[1][0], array[2][0], array[3][0]};
+  float box_y[4] = {array[0][1], array[1][1], array[2][1], array[3][1]};
+
+  int xmin = clamp(int(std::floorf(*(std::min_element(box_x, box_x + 4)))), 0,
+                   width - 1);
+  int xmax = clamp(int(std::ceilf(*(std::max_element(box_x, box_x + 4)))), 0,
+                   width - 1);
+  int ymin = clamp(int(std::floorf(*(std::min_element(box_y, box_y + 4)))), 0,
+                   height - 1);
+  int ymax = clamp(int(std::ceilf(*(std::max_element(box_y, box_y + 4)))), 0,
+                   height - 1);
+
+  cv::Mat mask;
+  mask = cv::Mat::zeros(ymax - ymin + 1, xmax - xmin + 1, CV_8UC1);
+
+  cv::Point root_point[4];
+  root_point[0] = cv::Point(int(array[0][0]) - xmin, int(array[0][1]) - ymin);
+  root_point[1] = cv::Point(int(array[1][0]) - xmin, int(array[1][1]) - ymin);
+  root_point[2] = cv::Point(int(array[2][0]) - xmin, int(array[2][1]) - ymin);
+  root_point[3] = cv::Point(int(array[3][0]) - xmin, int(array[3][1]) - ymin);
+  const cv::Point *ppt[1] = {root_point};
+  int npt[] = {4};
+  cv::fillPoly(mask, ppt, npt, 1, cv::Scalar(1));
+
+  cv::Mat croppedImg;
+  pred(cv::Rect(xmin, ymin, xmax - xmin + 1, ymax - ymin + 1))
+      .copyTo(croppedImg);
+
+  auto score = cv::mean(croppedImg, mask)[0];
+  return score;
+}
+
+std::vector<std::vector<std::vector<int>>>
+boxes_from_bitmap(const cv::Mat &pred, const cv::Mat &bitmap) {
+  const int min_size = 3;
+  const int max_candidates = 1000;
+  const float box_thresh = 0.5;
+
+  int width = bitmap.cols;
+  int height = bitmap.rows;
+
+  std::vector<std::vector<cv::Point>> contours;
+  std::vector<cv::Vec4i> hierarchy;
+
+  cv::findContours(bitmap, contours, hierarchy, cv::RETR_LIST,
+                   cv::CHAIN_APPROX_SIMPLE);
+
+  int num_contours =
+      contours.size() >= max_candidates ? max_candidates : contours.size();
+
+  std::vector<std::vector<std::vector<int>>> boxes;
+
+  for (int _i = 0; _i < num_contours; _i++) {
+    float ssid;
+    cv::RotatedRect box = cv::minAreaRect(contours[_i]);
+    auto array = get_mini_boxes(box, ssid);
+
+    auto box_for_unclip = array;
+    // end get_mini_box
+
+    if (ssid < min_size) {
+      continue;
+    }
+
+    float score;
+    score = box_score_fast(array, pred);
+    // end box_score_fast
+    if (score < box_thresh) {
+      continue;
+    }
+
+    // start for unclip
+    cv::RotatedRect points = unclip(box_for_unclip);
+    // end for unclip
+
+    cv::RotatedRect clipbox = points;
+    auto cliparray = get_mini_boxes(clipbox, ssid);
+
+    if (ssid < min_size + 2)
+      continue;
+
+    int dest_width = pred.cols;
+    int dest_height = pred.rows;
+    std::vector<std::vector<int>> intcliparray;
+
+    for (int num_pt = 0; num_pt < 4; num_pt++) {
+      std::vector<int> a{int(clampf(roundf(cliparray[num_pt][0] / float(width) *
+                                           float(dest_width)),
+                                    0, float(dest_width))),
+                         int(clampf(roundf(cliparray[num_pt][1] /
+                                           float(height) * float(dest_height)),
+                                    0, float(dest_height)))};
+      intcliparray.emplace_back(std::move(a));
+    }
+    boxes.emplace_back(std::move(intcliparray));
+
+  } // end for
+  return boxes;
+}
+
+int _max(int a, int b) { return a >= b ? a : b; }
+
+int _min(int a, int b) { return a >= b ? b : a; }
+
+std::vector<std::vector<std::vector<int>>>
+filter_tag_det_res(const std::vector<std::vector<std::vector<int>>> &o_boxes,
+                   float ratio_h, float ratio_w, const cv::Mat &srcimg) {
+  int oriimg_h = srcimg.rows;
+  int oriimg_w = srcimg.cols;
+  std::vector<std::vector<std::vector<int>>> boxes{o_boxes};
+  std::vector<std::vector<std::vector<int>>> root_points;
+  for (int n = 0; n < boxes.size(); n++) {
+    boxes[n] = order_points_clockwise(boxes[n]);
+    for (int m = 0; m < boxes[0].size(); m++) {
+      boxes[n][m][0] /= ratio_w;
+      boxes[n][m][1] /= ratio_h;
+
+      boxes[n][m][0] = int(_min(_max(boxes[n][m][0], 0), oriimg_w - 1));
+      boxes[n][m][1] = int(_min(_max(boxes[n][m][1], 0), oriimg_h - 1));
+    }
+  }
+
+  for (int n = 0; n < boxes.size(); n++) {
+    int rect_width, rect_height;
+    rect_width = int(sqrt(pow(boxes[n][0][0] - boxes[n][1][0], 2) +
+                          pow(boxes[n][0][1] - boxes[n][1][1], 2)));
+    rect_height = int(sqrt(pow(boxes[n][0][0] - boxes[n][3][0], 2) +
+                           pow(boxes[n][0][1] - boxes[n][3][1], 2)));
+    if (rect_width <= 10 || rect_height <= 10)
+      continue;
+    root_points.push_back(boxes[n]);
+  }
+  return root_points;
+}
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/ocr_db_post_process.h

+//
+// Created by fujiayi on 2020/7/2.
+//
+#pragma once
+#include <opencv2/opencv.hpp>
+#include <vector>
+
+std::vector<std::vector<std::vector<int>>>
+boxes_from_bitmap(const cv::Mat &pred, const cv::Mat &bitmap);
+
+std::vector<std::vector<std::vector<int>>>
+filter_tag_det_res(const std::vector<std::vector<std::vector<int>>> &o_boxes,
+                   float ratio_h, float ratio_w, const cv::Mat &srcimg);
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/ocr_ppredictor.cpp

+//
+// Created by fujiayi on 2020/7/1.
+//
+
+#include "ocr_ppredictor.h"
+#include "common.h"
+#include "ocr_cls_process.h"
+#include "ocr_crnn_process.h"
+#include "ocr_db_post_process.h"
+#include "preprocess.h"
+
+namespace ppredictor {
+
+OCR_PPredictor::OCR_PPredictor(const OCR_Config &config) : _config(config) {}
+
+int OCR_PPredictor::init(const std::string &det_model_content,
+                         const std::string &rec_model_content,
+                         const std::string &cls_model_content) {
+  _det_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.thread_num, NET_OCR, _config.mode});
+  _det_predictor->init_nb(det_model_content);
+
+  _rec_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.thread_num, NET_OCR_INTERNAL, _config.mode});
+  _rec_predictor->init_nb(rec_model_content);
+
+  _cls_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.thread_num, NET_OCR_INTERNAL, _config.mode});
+  _cls_predictor->init_nb(cls_model_content);
+  return RETURN_OK;
+}
+
+int OCR_PPredictor::init_from_file(const std::string &det_model_path,
+                                   const std::string &rec_model_path,
+                                   const std::string &cls_model_path) {
+  _det_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.thread_num, NET_OCR, _config.mode});
+  _det_predictor->init_from_file(det_model_path);
+
+  _rec_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.thread_num, NET_OCR_INTERNAL, _config.mode});
+  _rec_predictor->init_from_file(rec_model_path);
+
+  _cls_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.thread_num, NET_OCR_INTERNAL, _config.mode});
+  _cls_predictor->init_from_file(cls_model_path);
+  return RETURN_OK;
+}
+/**
+ * for debug use, show result of First Step
+ * @param filter_boxes
+ * @param boxes
+ * @param srcimg
+ */
+static void
+visual_img(const std::vector<std::vector<std::vector<int>>> &filter_boxes,
+           const std::vector<std::vector<std::vector<int>>> &boxes,
+           const cv::Mat &srcimg) {
+  // visualization
+  cv::Point rook_points[filter_boxes.size()][4];
+  for (int n = 0; n < filter_boxes.size(); n++) {
+    for (int m = 0; m < filter_boxes[0].size(); m++) {
+      rook_points[n][m] =
+          cv::Point(int(filter_boxes[n][m][0]), int(filter_boxes[n][m][1]));
+    }
+  }
+
+  cv::Mat img_vis;
+  srcimg.copyTo(img_vis);
+  for (int n = 0; n < boxes.size(); n++) {
+    const cv::Point *ppt[1] = {rook_points[n]};
+    int npt[] = {4};
+    cv::polylines(img_vis, ppt, npt, 1, 1, CV_RGB(0, 255, 0), 2, 8, 0);
+  }
+  // 调试用，自行替换需要修改的路径
+  cv::imwrite("/sdcard/1/vis.png", img_vis);
+}
+
+std::vector<OCRPredictResult>
+OCR_PPredictor::infer_ocr(const std::vector<int64_t> &dims,
+                          const float *input_data, int input_len, int net_flag,
+                          cv::Mat &origin) {
+  PredictorInput input = _det_predictor->get_first_input();
+  input.set_dims(dims);
+  input.set_data(input_data, input_len);
+  std::vector<PredictorOutput> results = _det_predictor->infer();
+  PredictorOutput &res = results.at(0);
+  std::vector<std::vector<std::vector<int>>> filtered_box = calc_filtered_boxes(
+      res.get_float_data(), res.get_size(), (int)dims[2], (int)dims[3], origin);
+  LOGI("Filter_box size %ld", filtered_box.size());
+  return infer_rec(filtered_box, origin);
+}
+
+std::vector<OCRPredictResult> OCR_PPredictor::infer_rec(
+    const std::vector<std::vector<std::vector<int>>> &boxes,
+    const cv::Mat &origin_img) {
+  std::vector<float> mean = {0.5f, 0.5f, 0.5f};
+  std::vector<float> scale = {1 / 0.5f, 1 / 0.5f, 1 / 0.5f};
+  std::vector<int64_t> dims = {1, 3, 0, 0};
+  std::vector<OCRPredictResult> ocr_results;
+
+  PredictorInput input = _rec_predictor->get_first_input();
+  for (auto bp = boxes.crbegin(); bp != boxes.crend(); ++bp) {
+    const std::vector<std::vector<int>> &box = *bp;
+    cv::Mat crop_img = get_rotate_crop_image(origin_img, box);
+    crop_img = infer_cls(crop_img);
+
+    float wh_ratio = float(crop_img.cols) / float(crop_img.rows);
+    cv::Mat input_image = crnn_resize_img(crop_img, wh_ratio);
+    input_image.convertTo(input_image, CV_32FC3, 1 / 255.0f);
+    const float *dimg = reinterpret_cast<const float *>(input_image.data);
+    int input_size = input_image.rows * input_image.cols;
+
+    dims[2] = input_image.rows;
+    dims[3] = input_image.cols;
+    input.set_dims(dims);
+
+    neon_mean_scale(dimg, input.get_mutable_float_data(), input_size, mean,
+                    scale);
+
+    std::vector<PredictorOutput> results = _rec_predictor->infer();
+    const float *predict_batch = results.at(0).get_float_data();
+    const std::vector<int64_t> predict_shape = results.at(0).get_shape();
+
+    OCRPredictResult res;
+
+    // ctc decode
+    int argmax_idx;
+    int last_index = 0;
+    float score = 0.f;
+    int count = 0;
+    float max_value = 0.0f;
+
+    for (int n = 0; n < predict_shape[1]; n++) {
+      argmax_idx = int(argmax(&predict_batch[n * predict_shape[2]],
+                              &predict_batch[(n + 1) * predict_shape[2]]));
+      max_value =
+          float(*std::max_element(&predict_batch[n * predict_shape[2]],
+                                  &predict_batch[(n + 1) * predict_shape[2]]));
+      if (argmax_idx > 0 && (!(n > 0 && argmax_idx == last_index))) {
+        score += max_value;
+        count += 1;
+        res.word_index.push_back(argmax_idx);
+      }
+      last_index = argmax_idx;
+    }
+    score /= count;
+    if (res.word_index.empty()) {
+      continue;
+    }
+    res.score = score;
+    res.points = box;
+    ocr_results.emplace_back(std::move(res));
+  }
+  LOGI("ocr_results finished %lu", ocr_results.size());
+  return ocr_results;
+}
+
+cv::Mat OCR_PPredictor::infer_cls(const cv::Mat &img, float thresh) {
+  std::vector<float> mean = {0.5f, 0.5f, 0.5f};
+  std::vector<float> scale = {1 / 0.5f, 1 / 0.5f, 1 / 0.5f};
+  std::vector<int64_t> dims = {1, 3, 0, 0};
+  std::vector<OCRPredictResult> ocr_results;
+
+  PredictorInput input = _cls_predictor->get_first_input();
+
+  cv::Mat input_image = cls_resize_img(img);
+  input_image.convertTo(input_image, CV_32FC3, 1 / 255.0f);
+  const float *dimg = reinterpret_cast<const float *>(input_image.data);
+  int input_size = input_image.rows * input_image.cols;
+
+  dims[2] = input_image.rows;
+  dims[3] = input_image.cols;
+  input.set_dims(dims);
+
+  neon_mean_scale(dimg, input.get_mutable_float_data(), input_size, mean,
+                  scale);
+
+  std::vector<PredictorOutput> results = _cls_predictor->infer();
+
+  const float *scores = results.at(0).get_float_data();
+  float score = 0;
+  int label = 0;
+  for (int64_t i = 0; i < results.at(0).get_size(); i++) {
+    LOGI("output scores [%f]", scores[i]);
+    if (scores[i] > score) {
+      score = scores[i];
+      label = i;
+    }
+  }
+  cv::Mat srcimg;
+  img.copyTo(srcimg);
+  if (label % 2 == 1 && score > thresh) {
+    cv::rotate(srcimg, srcimg, 1);
+  }
+  return srcimg;
+}
+
+std::vector<std::vector<std::vector<int>>>
+OCR_PPredictor::calc_filtered_boxes(const float *pred, int pred_size,
+                                    int output_height, int output_width,
+                                    const cv::Mat &origin) {
+  const double threshold = 0.3;
+  const double maxvalue = 1;
+
+  cv::Mat pred_map = cv::Mat::zeros(output_height, output_width, CV_32F);
+  memcpy(pred_map.data, pred, pred_size * sizeof(float));
+  cv::Mat cbuf_map;
+  pred_map.convertTo(cbuf_map, CV_8UC1);
+
+  cv::Mat bit_map;
+  cv::threshold(cbuf_map, bit_map, threshold, maxvalue, cv::THRESH_BINARY);
+
+  std::vector<std::vector<std::vector<int>>> boxes =
+      boxes_from_bitmap(pred_map, bit_map);
+  float ratio_h = output_height * 1.0f / origin.rows;
+  float ratio_w = output_width * 1.0f / origin.cols;
+  std::vector<std::vector<std::vector<int>>> filter_boxes =
+      filter_tag_det_res(boxes, ratio_h, ratio_w, origin);
+  return filter_boxes;
+}
+
+std::vector<int>
+OCR_PPredictor::postprocess_rec_word_index(const PredictorOutput &res) {
+  const int *rec_idx = res.get_int_data();
+  const std::vector<std::vector<uint64_t>> rec_idx_lod = res.get_lod();
+
+  std::vector<int> pred_idx;
+  for (int n = int(rec_idx_lod[0][0]); n < int(rec_idx_lod[0][1] * 2); n += 2) {
+    pred_idx.emplace_back(rec_idx[n]);
+  }
+  return pred_idx;
+}
+
+float OCR_PPredictor::postprocess_rec_score(const PredictorOutput &res) {
+  const float *predict_batch = res.get_float_data();
+  const std::vector<int64_t> predict_shape = res.get_shape();
+  const std::vector<std::vector<uint64_t>> predict_lod = res.get_lod();
+  int blank = predict_shape[1];
+  float score = 0.f;
+  int count = 0;
+  for (int n = predict_lod[0][0]; n < predict_lod[0][1] - 1; n++) {
+    int argmax_idx = argmax(predict_batch + n * predict_shape[1],
+                            predict_batch + (n + 1) * predict_shape[1]);
+    float max_value = predict_batch[n * predict_shape[1] + argmax_idx];
+    if (blank - 1 - argmax_idx > 1e-5) {
+      score += max_value;
+      count += 1;
+    }
+  }
+  if (count == 0) {
+    LOGE("calc score count 0");
+  } else {
+    score /= count;
+  }
+  LOGI("calc score: %f", score);
+  return score;
+}
+
+NET_TYPE OCR_PPredictor::get_net_flag() const { return NET_OCR; }
+}
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/ocr_ppredictor.h

+//
+// Created by fujiayi on 2020/7/1.
+//
+
+#pragma once
+
+#include "ppredictor.h"
+#include <opencv2/opencv.hpp>
+#include <paddle_api.h>
+#include <string>
+
+namespace ppredictor {
+
+/**
+ * Config
+ */
+struct OCR_Config {
+  int thread_num = 4; // Thread num
+  paddle::lite_api::PowerMode mode =
+      paddle::lite_api::LITE_POWER_HIGH; // PaddleLite Mode
+};
+
+/**
+ * PolyGone Result
+ */
+struct OCRPredictResult {
+  std::vector<int> word_index;
+  std::vector<std::vector<int>> points;
+  float score;
+};
+
+/**
+ * OCR there are 2 models
+ * 1. First model（det），select polygones to show where are the texts
+ * 2. crop from the origin images, use these polygones to infer
+ */
+class OCR_PPredictor : public PPredictor_Interface {
+public:
+  OCR_PPredictor(const OCR_Config &config);
+
+  virtual ~OCR_PPredictor() {}
+
+  /**
+   * 初始化二个模型的Predictor
+   * @param det_model_content
+   * @param rec_model_content
+   * @return
+   */
+  int init(const std::string &det_model_content,
+           const std::string &rec_model_content,
+           const std::string &cls_model_content);
+  int init_from_file(const std::string &det_model_path,
+                     const std::string &rec_model_path,
+                     const std::string &cls_model_path);
+  /**
+   * Return OCR result
+   * @param dims
+   * @param input_data
+   * @param input_len
+   * @param net_flag
+   * @param origin
+   * @return
+   */
+  virtual std::vector<OCRPredictResult>
+  infer_ocr(const std::vector<int64_t> &dims, const float *input_data,
+            int input_len, int net_flag, cv::Mat &origin);
+
+  virtual NET_TYPE get_net_flag() const;
+
+private:
+  /**
+   * calcul Polygone from the result image of first model
+   * @param pred
+   * @param output_height
+   * @param output_width
+   * @param origin
+   * @return
+   */
+  std::vector<std::vector<std::vector<int>>>
+  calc_filtered_boxes(const float *pred, int pred_size, int output_height,
+                      int output_width, const cv::Mat &origin);
+
+  /**
+   * infer for second model
+   *
+   * @param boxes
+   * @param origin
+   * @return
+   */
+  std::vector<OCRPredictResult>
+  infer_rec(const std::vector<std::vector<std::vector<int>>> &boxes,
+            const cv::Mat &origin);
+
+  /**
+  * infer for cls model
+  *
+  * @param boxes
+  * @param origin
+  * @return
+  */
+  cv::Mat infer_cls(const cv::Mat &origin, float thresh = 0.9);
+
+  /**
+   * Postprocess or sencod model to extract text
+   * @param res
+   * @return
+   */
+  std::vector<int> postprocess_rec_word_index(const PredictorOutput &res);
+
+  /**
+   * calculate confidence of second model text result
+   * @param res
+   * @return
+   */
+  float postprocess_rec_score(const PredictorOutput &res);
+
+  std::unique_ptr<PPredictor> _det_predictor;
+  std::unique_ptr<PPredictor> _rec_predictor;
+  std::unique_ptr<PPredictor> _cls_predictor;
+  OCR_Config _config;
+};
+}

deploy/android_demo/app/src/main/cpp/ppredictor.cpp

+#include "ppredictor.h"
+#include "common.h"
+
+namespace ppredictor {
+PPredictor::PPredictor(int thread_num, int net_flag,
+                       paddle::lite_api::PowerMode mode)
+    : _thread_num(thread_num), _net_flag(net_flag), _mode(mode) {}
+
+int PPredictor::init_nb(const std::string &model_content) {
+  paddle::lite_api::MobileConfig config;
+  config.set_model_from_buffer(model_content);
+  return _init(config);
+}
+
+int PPredictor::init_from_file(const std::string &model_content) {
+  paddle::lite_api::MobileConfig config;
+  config.set_model_from_file(model_content);
+  return _init(config);
+}
+
+template <typename ConfigT> int PPredictor::_init(ConfigT &config) {
+  config.set_threads(_thread_num);
+  config.set_power_mode(_mode);
+  _predictor = paddle::lite_api::CreatePaddlePredictor(config);
+  LOGI("paddle instance created");
+  return RETURN_OK;
+}
+
+PredictorInput PPredictor::get_input(int index) {
+  PredictorInput input{_predictor->GetInput(index), index, _net_flag};
+  _is_input_get = true;
+  return input;
+}
+
+std::vector<PredictorInput> PPredictor::get_inputs(int num) {
+  std::vector<PredictorInput> results;
+  for (int i = 0; i < num; i++) {
+    results.emplace_back(get_input(i));
+  }
+  return results;
+}
+
+PredictorInput PPredictor::get_first_input() { return get_input(0); }
+
+std::vector<PredictorOutput> PPredictor::infer() {
+  LOGI("infer Run start %d", _net_flag);
+  std::vector<PredictorOutput> results;
+  if (!_is_input_get) {
+    return results;
+  }
+  _predictor->Run();
+  LOGI("infer Run end");
+
+  for (int i = 0; i < _predictor->GetOutputNames().size(); i++) {
+    std::unique_ptr<const paddle::lite_api::Tensor> output_tensor =
+        _predictor->GetOutput(i);
+    LOGI("output tensor[%d] size %ld", i, product(output_tensor->shape()));
+    PredictorOutput result{std::move(output_tensor), i, _net_flag};
+    results.emplace_back(std::move(result));
+  }
+  return results;
+}
+
+NET_TYPE PPredictor::get_net_flag() const { return (NET_TYPE)_net_flag; }
+}
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/ppredictor.h

+#pragma once
+
+#include "paddle_api.h"
+#include "predictor_input.h"
+#include "predictor_output.h"
+
+namespace ppredictor {
+
+/**
+ * PaddleLite Preditor Common Interface
+ */
+class PPredictor_Interface {
+public:
+  virtual ~PPredictor_Interface() {}
+
+  virtual NET_TYPE get_net_flag() const = 0;
+};
+
+/**
+ * Common Predictor
+ */
+class PPredictor : public PPredictor_Interface {
+public:
+  PPredictor(
+      int thread_num, int net_flag = 0,
+      paddle::lite_api::PowerMode mode = paddle::lite_api::LITE_POWER_HIGH);
+
+  virtual ~PPredictor() {}
+
+  /**
+   * init paddlitelite opt model，nb format ，or use ini_paddle
+   * @param model_content
+   * @return 0
+   */
+  virtual int init_nb(const std::string &model_content);
+
+  virtual int init_from_file(const std::string &model_content);
+
+  std::vector<PredictorOutput> infer();
+
+  std::shared_ptr<paddle::lite_api::PaddlePredictor> get_predictor() {
+    return _predictor;
+  }
+
+  virtual std::vector<PredictorInput> get_inputs(int num);
+
+  virtual PredictorInput get_input(int index);
+
+  virtual PredictorInput get_first_input();
+
+  virtual NET_TYPE get_net_flag() const;
+
+protected:
+  template <typename ConfigT> int _init(ConfigT &config);
+
+private:
+  int _thread_num;
+  paddle::lite_api::PowerMode _mode;
+  std::shared_ptr<paddle::lite_api::PaddlePredictor> _predictor;
+  bool _is_input_get = false;
+  int _net_flag;
+};
+}

deploy/android_demo/app/src/main/cpp/predictor_input.cpp

+#include "predictor_input.h"
+
+namespace ppredictor {
+
+void PredictorInput::set_dims(std::vector<int64_t> dims) {
+  // yolov3
+  if (_net_flag == 101 && _index == 1) {
+    _tensor->Resize({1, 2});
+    _tensor->mutable_data<int>()[0] = (int)dims.at(2);
+    _tensor->mutable_data<int>()[1] = (int)dims.at(3);
+  } else {
+    _tensor->Resize(dims);
+  }
+  _is_dims_set = true;
+}
+
+float *PredictorInput::get_mutable_float_data() {
+  if (!_is_dims_set) {
+    LOGE("PredictorInput::set_dims is not called");
+  }
+  return _tensor->mutable_data<float>();
+}
+
+void PredictorInput::set_data(const float *input_data, int input_float_len) {
+  float *input_raw_data = get_mutable_float_data();
+  memcpy(input_raw_data, input_data, input_float_len * sizeof(float));
+}
+}
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/predictor_input.h

+#pragma once
+
+#include "common.h"
+#include <paddle_api.h>
+#include <vector>
+
+namespace ppredictor {
+class PredictorInput {
+public:
+  PredictorInput(std::unique_ptr<paddle::lite_api::Tensor> &&tensor, int index,
+                 int net_flag)
+      : _tensor(std::move(tensor)), _index(index), _net_flag(net_flag) {}
+
+  void set_dims(std::vector<int64_t> dims);
+
+  float *get_mutable_float_data();
+
+  void set_data(const float *input_data, int input_float_len);
+
+private:
+  std::unique_ptr<paddle::lite_api::Tensor> _tensor;
+  bool _is_dims_set = false;
+  int _index;
+  int _net_flag;
+};
+}

deploy/android_demo/app/src/main/cpp/predictor_output.cpp

+#include "predictor_output.h"
+namespace ppredictor {
+const float *PredictorOutput::get_float_data() const {
+  return _tensor->data<float>();
+}
+
+const int *PredictorOutput::get_int_data() const {
+  return _tensor->data<int>();
+}
+
+const std::vector<std::vector<uint64_t>> PredictorOutput::get_lod() const {
+  return _tensor->lod();
+}
+
+int64_t PredictorOutput::get_size() const {
+  if (_net_flag == NET_OCR) {
+    return _tensor->shape().at(2) * _tensor->shape().at(3);
+  } else {
+    return product(_tensor->shape());
+  }
+}
+
+const std::vector<int64_t> PredictorOutput::get_shape() const {
+  return _tensor->shape();
+}
+}
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/predictor_output.h

+#pragma once
+
+#include "common.h"
+#include <paddle_api.h>
+#include <vector>
+
+namespace ppredictor {
+class PredictorOutput {
+public:
+  PredictorOutput() {}
+  PredictorOutput(std::unique_ptr<const paddle::lite_api::Tensor> &&tensor,
+                  int index, int net_flag)
+      : _tensor(std::move(tensor)), _index(index), _net_flag(net_flag) {}
+
+  const float *get_float_data() const;
+  const int *get_int_data() const;
+  int64_t get_size() const;
+  const std::vector<std::vector<uint64_t>> get_lod() const;
+  const std::vector<int64_t> get_shape() const;
+
+  std::vector<float> data;    // return float, or use data_int
+  std::vector<int> data_int;  // several layers return int ，or use data
+  std::vector<int64_t> shape; // PaddleLite output shape
+  std::vector<std::vector<uint64_t>> lod; // PaddleLite output lod
+
+private:
+  std::unique_ptr<const paddle::lite_api::Tensor> _tensor;
+  int _index;
+  int _net_flag;
+};
+}

deploy/android_demo/app/src/main/cpp/preprocess.cpp

+#include "preprocess.h"
+#include <android/bitmap.h>
+
+cv::Mat bitmap_to_cv_mat(JNIEnv *env, jobject bitmap) {
+  AndroidBitmapInfo info;
+  int result = AndroidBitmap_getInfo(env, bitmap, &info);
+  if (result != ANDROID_BITMAP_RESULT_SUCCESS) {
+    LOGE("AndroidBitmap_getInfo failed, result: %d", result);
+    return cv::Mat{};
+  }
+  if (info.format != ANDROID_BITMAP_FORMAT_RGBA_8888) {
+    LOGE("Bitmap format is not RGBA_8888 !");
+    return cv::Mat{};
+  }
+  unsigned char *srcData = NULL;
+  AndroidBitmap_lockPixels(env, bitmap, (void **)&srcData);
+  cv::Mat mat = cv::Mat::zeros(info.height, info.width, CV_8UC4);
+  memcpy(mat.data, srcData, info.height * info.width * 4);
+  AndroidBitmap_unlockPixels(env, bitmap);
+  cv::cvtColor(mat, mat, cv::COLOR_RGBA2BGR);
+  /**
+  if (!cv::imwrite("/sdcard/1/copy.jpg", mat)){
+      LOGE("Write image failed " );
+  }
+   */
+
+  return mat;
+}
+
+cv::Mat resize_img(const cv::Mat &img, int height, int width) {
+  if (img.rows == height && img.cols == width) {
+    return img;
+  }
+  cv::Mat new_img;
+  cv::resize(img, new_img, cv::Size(height, width));
+  return new_img;
+}
+
+// fill tensor with mean and scale and trans layout: nhwc -> nchw, neon speed up
+void neon_mean_scale(const float *din, float *dout, int size,
+                     const std::vector<float> &mean,
+                     const std::vector<float> &scale) {
+  if (mean.size() != 3 || scale.size() != 3) {
+    LOGE("[ERROR] mean or scale size must equal to 3");
+    return;
+  }
+
+  float32x4_t vmean0 = vdupq_n_f32(mean[0]);
+  float32x4_t vmean1 = vdupq_n_f32(mean[1]);
+  float32x4_t vmean2 = vdupq_n_f32(mean[2]);
+  float32x4_t vscale0 = vdupq_n_f32(scale[0]);
+  float32x4_t vscale1 = vdupq_n_f32(scale[1]);
+  float32x4_t vscale2 = vdupq_n_f32(scale[2]);
+
+  float *dout_c0 = dout;
+  float *dout_c1 = dout + size;
+  float *dout_c2 = dout + size * 2;
+
+  int i = 0;
+  for (; i < size - 3; i += 4) {
+    float32x4x3_t vin3 = vld3q_f32(din);
+    float32x4_t vsub0 = vsubq_f32(vin3.val[0], vmean0);
+    float32x4_t vsub1 = vsubq_f32(vin3.val[1], vmean1);
+    float32x4_t vsub2 = vsubq_f32(vin3.val[2], vmean2);
+    float32x4_t vs0 = vmulq_f32(vsub0, vscale0);
+    float32x4_t vs1 = vmulq_f32(vsub1, vscale1);
+    float32x4_t vs2 = vmulq_f32(vsub2, vscale2);
+    vst1q_f32(dout_c0, vs0);
+    vst1q_f32(dout_c1, vs1);
+    vst1q_f32(dout_c2, vs2);
+
+    din += 12;
+    dout_c0 += 4;
+    dout_c1 += 4;
+    dout_c2 += 4;
+  }
+  for (; i < size; i++) {
+    *(dout_c0++) = (*(din++) - mean[0]) * scale[0];
+    *(dout_c1++) = (*(din++) - mean[1]) * scale[1];
+    *(dout_c2++) = (*(din++) - mean[2]) * scale[2];
+  }
+}
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/preprocess.h

+#pragma once
+
+#include "common.h"
+#include <jni.h>
+#include <opencv2/opencv.hpp>
+cv::Mat bitmap_to_cv_mat(JNIEnv *env, jobject bitmap);
+
+cv::Mat resize_img(const cv::Mat &img, int height, int width);
+
+void neon_mean_scale(const float *din, float *dout, int size,
+                     const std::vector<float> &mean,
+                     const std::vector<float> &scale);

deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/AppCompatPreferenceActivity.java

+/*
+ * Copyright (C) 2014 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package com.baidu.paddle.lite.demo.ocr;
+
+import android.content.res.Configuration;
+import android.os.Bundle;
+import android.preference.PreferenceActivity;
+import android.view.MenuInflater;
+import android.view.View;
+import android.view.ViewGroup;
+
+import androidx.annotation.LayoutRes;
+import androidx.annotation.Nullable;
+import androidx.appcompat.app.ActionBar;
+import androidx.appcompat.app.AppCompatDelegate;
+import androidx.appcompat.widget.Toolbar;
+
+/**
+ * A {@link PreferenceActivity} which implements and proxies the necessary calls
+ * to be used with AppCompat.
+ * <p>
+ * This technique can be used with an {@link android.app.Activity} class, not just
+ * {@link PreferenceActivity}.
+ */
+public abstract class AppCompatPreferenceActivity extends PreferenceActivity {
+    private AppCompatDelegate mDelegate;
+
+    @Override
+    protected void onCreate(Bundle savedInstanceState) {
+        getDelegate().installViewFactory();
+        getDelegate().onCreate(savedInstanceState);
+        super.onCreate(savedInstanceState);
+    }
+
+    @Override
+    protected void onPostCreate(Bundle savedInstanceState) {
+        super.onPostCreate(savedInstanceState);
+        getDelegate().onPostCreate(savedInstanceState);
+    }
+
+    public ActionBar getSupportActionBar() {
+        return getDelegate().getSupportActionBar();
+    }
+
+    public void setSupportActionBar(@Nullable Toolbar toolbar) {
+        getDelegate().setSupportActionBar(toolbar);
+    }
+
+    @Override
+    public MenuInflater getMenuInflater() {
+        return getDelegate().getMenuInflater();
+    }
+
+    @Override
+    public void setContentView(@LayoutRes int layoutResID) {
+        getDelegate().setContentView(layoutResID);
+    }
+
+    @Override
+    public void setContentView(View view) {
+        getDelegate().setContentView(view);
+    }
+
+    @Override
+    public void setContentView(View view, ViewGroup.LayoutParams params) {
+        getDelegate().setContentView(view, params);
+    }
+
+    @Override
+    public void addContentView(View view, ViewGroup.LayoutParams params) {
+        getDelegate().addContentView(view, params);
+    }
+
+    @Override
+    protected void onPostResume() {
+        super.onPostResume();
+        getDelegate().onPostResume();
+    }
+
+    @Override
+    protected void onTitleChanged(CharSequence title, int color) {
+        super.onTitleChanged(title, color);
+        getDelegate().setTitle(title);
+    }
+
+    @Override
+    public void onConfigurationChanged(Configuration newConfig) {
+        super.onConfigurationChanged(newConfig);
+        getDelegate().onConfigurationChanged(newConfig);
+    }
+
+    @Override
+    protected void onStop() {
+        super.onStop();
+        getDelegate().onStop();
+    }
+
+    @Override
+    protected void onDestroy() {
+        super.onDestroy();
+        getDelegate().onDestroy();
+    }
+
+    public void invalidateOptionsMenu() {
+        getDelegate().invalidateOptionsMenu();
+    }
+
+    private AppCompatDelegate getDelegate() {
+        if (mDelegate == null) {
+            mDelegate = AppCompatDelegate.create(this, null);
+        }
+        return mDelegate;
+    }
+}

deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/MiniActivity.java

+package com.baidu.paddle.lite.demo.ocr;
+
+import android.graphics.Bitmap;
+import android.graphics.BitmapFactory;
+import android.os.Build;
+import android.os.Bundle;
+import android.os.Handler;
+import android.os.HandlerThread;
+import android.os.Message;
+import android.util.Log;
+import android.view.View;
+import android.widget.Button;
+import android.widget.ImageView;
+import android.widget.TextView;
+import android.widget.Toast;
+
+import androidx.appcompat.app.AppCompatActivity;
+
+import java.io.IOException;
+import java.io.InputStream;
+
+public class MiniActivity extends AppCompatActivity {
+
+
+    public static final int REQUEST_LOAD_MODEL = 0;
+    public static final int REQUEST_RUN_MODEL = 1;
+    public static final int REQUEST_UNLOAD_MODEL = 2;
+    public static final int RESPONSE_LOAD_MODEL_SUCCESSED = 0;
+    public static final int RESPONSE_LOAD_MODEL_FAILED = 1;
+    public static final int RESPONSE_RUN_MODEL_SUCCESSED = 2;
+    public static final int RESPONSE_RUN_MODEL_FAILED = 3;
+
+    private static final String TAG = "MiniActivity";
+
+    protected Handler receiver = null; // Receive messages from worker thread
+    protected Handler sender = null; // Send command to worker thread
+    protected HandlerThread worker = null; // Worker thread to load&run model
+    protected volatile Predictor predictor = null;
+
+    private String assetModelDirPath = "models/ocr_v2_for_cpu";
+    private String assetlabelFilePath = "labels/ppocr_keys_v1.txt";
+
+    private Button button;
+    private ImageView imageView; // image result
+    private TextView textView; // text result
+
+    @Override
+    protected void onCreate(Bundle savedInstanceState) {
+        super.onCreate(savedInstanceState);
+        setContentView(R.layout.activity_mini);
+
+        Log.i(TAG, "SHOW in Logcat");
+
+        // Prepare the worker thread for mode loading and inference
+        worker = new HandlerThread("Predictor Worker");
+        worker.start();
+        sender = new Handler(worker.getLooper()) {
+            public void handleMessage(Message msg) {
+                switch (msg.what) {
+                    case REQUEST_LOAD_MODEL:
+                        // Load model and reload test image
+                        if (!onLoadModel()) {
+                            runOnUiThread(new Runnable() {
+                                @Override
+                                public void run() {
+                                    Toast.makeText(MiniActivity.this, "Load model failed!", Toast.LENGTH_SHORT).show();
+                                }
+                            });
+                        }
+                        break;
+                    case REQUEST_RUN_MODEL:
+                        // Run model if model is loaded
+                        final boolean isSuccessed = onRunModel();
+                        runOnUiThread(new Runnable() {
+                            @Override
+                            public void run() {
+                                if (isSuccessed){
+                                    onRunModelSuccessed();
+                                }else{
+                                    Toast.makeText(MiniActivity.this, "Run model failed!", Toast.LENGTH_SHORT).show();
+                                }
+                            }
+                        });
+                        break;
+                }
+            }
+        };
+        sender.sendEmptyMessage(REQUEST_LOAD_MODEL); // corresponding to REQUEST_LOAD_MODEL， to call onLoadModel()
+
+        imageView = findViewById(R.id.imageView);
+        textView = findViewById(R.id.sample_text);
+        button = findViewById(R.id.button);
+        button.setOnClickListener(new View.OnClickListener() {
+            @Override
+            public void onClick(View v) {
+                sender.sendEmptyMessage(REQUEST_RUN_MODEL);
+            }
+        });
+
+
+    }
+
+    @Override
+    protected void onDestroy() {
+        onUnloadModel();
+        if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.JELLY_BEAN_MR2) {
+            worker.quitSafely();
+        } else {
+            worker.quit();
+        }
+        super.onDestroy();
+    }
+
+    /**
+     * call in onCreate, model init
+     *
+     * @return
+     */
+    private boolean onLoadModel() {
+        if (predictor == null) {
+            predictor = new Predictor();
+        }
+        return predictor.init(this, assetModelDirPath, assetlabelFilePath);
+    }
+
+    /**
+     * init engine
+     * call in onCreate
+     *
+     * @return
+     */
+    private boolean onRunModel() {
+        try {
+            String assetImagePath = "images/0.jpg";
+            InputStream imageStream = getAssets().open(assetImagePath);
+            Bitmap image = BitmapFactory.decodeStream(imageStream);
+            // Input is Bitmap
+            predictor.setInputImage(image);
+            return predictor.isLoaded() && predictor.runModel();
+        } catch (IOException e) {
+            e.printStackTrace();
+            return false;
+        }
+    }
+
+    private void onRunModelSuccessed() {
+        Log.i(TAG, "onRunModelSuccessed");
+        textView.setText(predictor.outputResult);
+        imageView.setImageBitmap(predictor.outputImage);
+    }
+
+    private void onUnloadModel() {
+        if (predictor != null) {
+            predictor.releaseModel();
+        }
+    }
+}

deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/OCRPredictorNative.java

+package com.baidu.paddle.lite.demo.ocr;
+
+import android.graphics.Bitmap;
+import android.util.Log;
+
+import java.util.ArrayList;
+import java.util.concurrent.atomic.AtomicBoolean;
+
+public class OCRPredictorNative {
+
+    private static final AtomicBoolean isSOLoaded = new AtomicBoolean();
+
+    public static void loadLibrary() throws RuntimeException {
+        if (!isSOLoaded.get() && isSOLoaded.compareAndSet(false, true)) {
+            try {
+                System.loadLibrary("Native");
+            } catch (Throwable e) {
+                RuntimeException exception = new RuntimeException(
+                        "Load libNative.so failed, please check it exists in apk file.", e);
+                throw exception;
+            }
+        }
+    }
+
+    private Config config;
+
+    private long nativePointer = 0;
+
+    public OCRPredictorNative(Config config) {
+        this.config = config;
+        loadLibrary();
+        nativePointer = init(config.detModelFilename, config.recModelFilename,config.clsModelFilename,
+                config.cpuThreadNum, config.cpuPower);
+        Log.i("OCRPredictorNative", "load success " + nativePointer);
+
+    }
+
+
+    public ArrayList<OcrResultModel> runImage(float[] inputData, int width, int height, int channels, Bitmap originalImage) {
+        Log.i("OCRPredictorNative", "begin to run image " + inputData.length + " " + width + " " + height);
+        float[] dims = new float[]{1, channels, height, width};
+        float[] rawResults = forward(nativePointer, inputData, dims, originalImage);
+        ArrayList<OcrResultModel> results = postprocess(rawResults);
+        return results;
+    }
+
+    public static class Config {
+        public int cpuThreadNum;
+        public String cpuPower;
+        public String detModelFilename;
+        public String recModelFilename;
+        public String clsModelFilename;
+
+    }
+
+    public void destory(){
+        if (nativePointer > 0) {
+            release(nativePointer);
+            nativePointer = 0;
+        }
+    }
+
+    protected native long init(String detModelPath, String recModelPath,String clsModelPath, int threadNum, String cpuMode);
+
+    protected native float[] forward(long pointer, float[] buf, float[] ddims, Bitmap originalImage);
+
+    protected native void release(long pointer);
+
+    private ArrayList<OcrResultModel> postprocess(float[] raw) {
+        ArrayList<OcrResultModel> results = new ArrayList<OcrResultModel>();
+        int begin = 0;
+
+        while (begin < raw.length) {
+            int point_num = Math.round(raw[begin]);
+            int word_num = Math.round(raw[begin + 1]);
+            OcrResultModel model = parse(raw, begin + 2, point_num, word_num);
+            begin += 2 + 1 + point_num * 2 + word_num;
+            results.add(model);
+        }
+
+        return results;
+    }
+
+    private OcrResultModel parse(float[] raw, int begin, int pointNum, int wordNum) {
+        int current = begin;
+        OcrResultModel model = new OcrResultModel();
+        model.setConfidence(raw[current]);
+        current++;
+        for (int i = 0; i < pointNum; i++) {
+            model.addPoints(Math.round(raw[current + i * 2]), Math.round(raw[current + i * 2 + 1]));
+        }
+        current += (pointNum * 2);
+        for (int i = 0; i < wordNum; i++) {
+            int index = Math.round(raw[current + i]);
+            model.addWordIndex(index);
+        }
+        Log.i("OCRPredictorNative", "word finished " + wordNum);
+        return model;
+    }
+
+
+}

deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/OcrResultModel.java

+package com.baidu.paddle.lite.demo.ocr;
+
+import android.graphics.Point;
+
+import java.util.ArrayList;
+import java.util.List;
+
+public class OcrResultModel {
+    private List<Point> points;
+    private List<Integer> wordIndex;
+    private String label;
+    private float confidence;
+
+    public OcrResultModel() {
+        super();
+        points = new ArrayList<>();
+        wordIndex = new ArrayList<>();
+    }
+
+    public void addPoints(int x, int y) {
+        Point point = new Point(x, y);
+        points.add(point);
+    }
+
+    public void addWordIndex(int index) {
+        wordIndex.add(index);
+    }
+
+    public List<Point> getPoints() {
+        return points;
+    }
+
+    public List<Integer> getWordIndex() {
+        return wordIndex;
+    }
+
+    public String getLabel() {
+        return label;
+    }
+
+    public void setLabel(String label) {
+        this.label = label;
+    }
+
+    public float getConfidence() {
+        return confidence;
+    }
+
+    public void setConfidence(float confidence) {
+        this.confidence = confidence;
+    }
+}

deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/Predictor.java

+package com.baidu.paddle.lite.demo.ocr;
+
+import android.content.Context;
+import android.graphics.Bitmap;
+import android.graphics.Canvas;
+import android.graphics.Color;
+import android.graphics.Paint;
+import android.graphics.Path;
+import android.graphics.Point;
+import android.util.Log;
+
+import java.io.File;
+import java.io.InputStream;
+import java.util.ArrayList;
+import java.util.Date;
+import java.util.List;
+import java.util.Vector;
+
+import static android.graphics.Color.*;
+
+public class Predictor {
+    private static final String TAG = Predictor.class.getSimpleName();
+    public boolean isLoaded = false;
+    public int warmupIterNum = 1;
+    public int inferIterNum = 1;
+    public int cpuThreadNum = 4;
+    public String cpuPowerMode = "LITE_POWER_HIGH";
+    public String modelPath = "";
+    public String modelName = "";
+    protected OCRPredictorNative paddlePredictor = null;
+    protected float inferenceTime = 0;
+    // Only for object detection
+    protected Vector<String> wordLabels = new Vector<String>();
+    protected String inputColorFormat = "BGR";
+    protected long[] inputShape = new long[]{1, 3, 960};
+    protected float[] inputMean = new float[]{0.485f, 0.456f, 0.406f};
+    protected float[] inputStd = new float[]{1.0f / 0.229f, 1.0f / 0.224f, 1.0f / 0.225f};
+    protected float scoreThreshold = 0.1f;
+    protected Bitmap inputImage = null;
+    protected Bitmap outputImage = null;
+    protected volatile String outputResult = "";
+    protected float preprocessTime = 0;
+    protected float postprocessTime = 0;
+
+
+    public Predictor() {
+    }
+
+    public boolean init(Context appCtx, String modelPath, String labelPath) {
+        isLoaded = loadModel(appCtx, modelPath, cpuThreadNum, cpuPowerMode);
+        if (!isLoaded) {
+            return false;
+        }
+        isLoaded = loadLabel(appCtx, labelPath);
+        return isLoaded;
+    }
+
+
+    public boolean init(Context appCtx, String modelPath, String labelPath, int cpuThreadNum, String cpuPowerMode,
+                        String inputColorFormat,
+                        long[] inputShape, float[] inputMean,
+                        float[] inputStd, float scoreThreshold) {
+        if (inputShape.length != 3) {
+            Log.e(TAG, "Size of input shape should be: 3");
+            return false;
+        }
+        if (inputMean.length != inputShape[1]) {
+            Log.e(TAG, "Size of input mean should be: " + Long.toString(inputShape[1]));
+            return false;
+        }
+        if (inputStd.length != inputShape[1]) {
+            Log.e(TAG, "Size of input std should be: " + Long.toString(inputShape[1]));
+            return false;
+        }
+        if (inputShape[0] != 1) {
+            Log.e(TAG, "Only one batch is supported in the image classification demo, you can use any batch size in " +
+                    "your Apps!");
+            return false;
+        }
+        if (inputShape[1] != 1 && inputShape[1] != 3) {
+            Log.e(TAG, "Only one/three channels are supported in the image classification demo, you can use any " +
+                    "channel size in your Apps!");
+            return false;
+        }
+        if (!inputColorFormat.equalsIgnoreCase("BGR")) {
+            Log.e(TAG, "Only  BGR color format is supported.");
+            return false;
+        }
+        boolean isLoaded = init(appCtx, modelPath, labelPath);
+        if (!isLoaded) {
+            return false;
+        }
+        this.inputColorFormat = inputColorFormat;
+        this.inputShape = inputShape;
+        this.inputMean = inputMean;
+        this.inputStd = inputStd;
+        this.scoreThreshold = scoreThreshold;
+        return true;
+    }
+
+    protected boolean loadModel(Context appCtx, String modelPath, int cpuThreadNum, String cpuPowerMode) {
+        // Release model if exists
+        releaseModel();
+
+        // Load model
+        if (modelPath.isEmpty()) {
+            return false;
+        }
+        String realPath = modelPath;
+        if (!modelPath.substring(0, 1).equals("/")) {
+            // Read model files from custom path if the first character of mode path is '/'
+            // otherwise copy model to cache from assets
+            realPath = appCtx.getCacheDir() + "/" + modelPath;
+            Utils.copyDirectoryFromAssets(appCtx, modelPath, realPath);
+        }
+        if (realPath.isEmpty()) {
+            return false;
+        }
+
+        OCRPredictorNative.Config config = new OCRPredictorNative.Config();
+        config.cpuThreadNum = cpuThreadNum;
+        config.detModelFilename = realPath + File.separator + "ch_ppocr_mobile_v2.0_det_opt.nb";
+        config.recModelFilename = realPath + File.separator + "ch_ppocr_mobile_v2.0_rec_opt.nb";
+        config.clsModelFilename = realPath + File.separator + "ch_ppocr_mobile_v2.0_cls_opt.nb";
+        Log.e("Predictor", "model path" + config.detModelFilename + " ; " + config.recModelFilename + ";" + config.clsModelFilename);
+        config.cpuPower = cpuPowerMode;
+        paddlePredictor = new OCRPredictorNative(config);
+
+        this.cpuThreadNum = cpuThreadNum;
+        this.cpuPowerMode = cpuPowerMode;
+        this.modelPath = realPath;
+        this.modelName = realPath.substring(realPath.lastIndexOf("/") + 1);
+        return true;
+    }
+
+    public void releaseModel() {
+        if (paddlePredictor != null) {
+            paddlePredictor.destory();
+            paddlePredictor = null;
+        }
+        isLoaded = false;
+        cpuThreadNum = 1;
+        cpuPowerMode = "LITE_POWER_HIGH";
+        modelPath = "";
+        modelName = "";
+    }
+
+    protected boolean loadLabel(Context appCtx, String labelPath) {
+        wordLabels.clear();
+        wordLabels.add("black");
+        // Load word labels from file
+        try {
+            InputStream assetsInputStream = appCtx.getAssets().open(labelPath);
+            int available = assetsInputStream.available();
+            byte[] lines = new byte[available];
+            assetsInputStream.read(lines);
+            assetsInputStream.close();
+            String words = new String(lines);
+            String[] contents = words.split("\n");
+            for (String content : contents) {
+                wordLabels.add(content);
+            }
+            Log.i(TAG, "Word label size: " + wordLabels.size());
+        } catch (Exception e) {
+            Log.e(TAG, e.getMessage());
+            return false;
+        }
+        return true;
+    }
+
+
+    public boolean runModel() {
+        if (inputImage == null || !isLoaded()) {
+            return false;
+        }
+
+        // Pre-process image, and feed input tensor with pre-processed data
+
+        Bitmap scaleImage = Utils.resizeWithStep(inputImage, Long.valueOf(inputShape[2]).intValue(), 32);
+
+        Date start = new Date();
+        int channels = (int) inputShape[1];
+        int width = scaleImage.getWidth();
+        int height = scaleImage.getHeight();
+        float[] inputData = new float[channels * width * height];
+        if (channels == 3) {
+            int[] channelIdx = null;
+            if (inputColorFormat.equalsIgnoreCase("RGB")) {
+                channelIdx = new int[]{0, 1, 2};
+            } else if (inputColorFormat.equalsIgnoreCase("BGR")) {
+                channelIdx = new int[]{2, 1, 0};
+            } else {
+                Log.i(TAG, "Unknown color format " + inputColorFormat + ", only RGB and BGR color format is " +
+                        "supported!");
+                return false;
+            }
+            int[] channelStride = new int[]{width * height, width * height * 2};
+            int p = scaleImage.getPixel(scaleImage.getWidth() - 1, scaleImage.getHeight() - 1);
+            for (int y = 0; y < height; y++) {
+                for (int x = 0; x < width; x++) {
+                    int color = scaleImage.getPixel(x, y);
+                    float[] rgb = new float[]{(float) red(color) / 255.0f, (float) green(color) / 255.0f,
+                            (float) blue(color) / 255.0f};
+                    inputData[y * width + x] = (rgb[channelIdx[0]] - inputMean[0]) / inputStd[0];
+                    inputData[y * width + x + channelStride[0]] = (rgb[channelIdx[1]] - inputMean[1]) / inputStd[1];
+                    inputData[y * width + x + channelStride[1]] = (rgb[channelIdx[2]] - inputMean[2]) / inputStd[2];
+
+                }
+            }
+        } else if (channels == 1) {
+            for (int y = 0; y < height; y++) {
+                for (int x = 0; x < width; x++) {
+                    int color = inputImage.getPixel(x, y);
+                    float gray = (float) (red(color) + green(color) + blue(color)) / 3.0f / 255.0f;
+                    inputData[y * width + x] = (gray - inputMean[0]) / inputStd[0];
+                }
+            }
+        } else {
+            Log.i(TAG, "Unsupported channel size " + Integer.toString(channels) + ",  only channel 1 and 3 is " +
+                    "supported!");
+            return false;
+        }
+        float[] pixels = inputData;
+        Log.i(TAG, "pixels " + pixels[0] + " " + pixels[1] + " " + pixels[2] + " " + pixels[3]
+                + " " + pixels[pixels.length / 2] + " " + pixels[pixels.length / 2 + 1] + " " + pixels[pixels.length - 2] + " " + pixels[pixels.length - 1]);
+        Date end = new Date();
+        preprocessTime = (float) (end.getTime() - start.getTime());
+
+        // Warm up
+        for (int i = 0; i < warmupIterNum; i++) {
+            paddlePredictor.runImage(inputData, width, height, channels, inputImage);
+        }
+        warmupIterNum = 0; // do not need warm
+        // Run inference
+        start = new Date();
+        ArrayList<OcrResultModel> results = paddlePredictor.runImage(inputData, width, height, channels, inputImage);
+        end = new Date();
+        inferenceTime = (end.getTime() - start.getTime()) / (float) inferIterNum;
+
+        results = postprocess(results);
+        Log.i(TAG, "[stat] Preprocess Time: " + preprocessTime
+                + " ; Inference Time: " + inferenceTime + " ;Box Size " + results.size());
+        drawResults(results);
+
+        return true;
+    }
+
+
+    public boolean isLoaded() {
+        return paddlePredictor != null && isLoaded;
+    }
+
+    public String modelPath() {
+        return modelPath;
+    }
+
+    public String modelName() {
+        return modelName;
+    }
+
+    public int cpuThreadNum() {
+        return cpuThreadNum;
+    }
+
+    public String cpuPowerMode() {
+        return cpuPowerMode;
+    }
+
+    public float inferenceTime() {
+        return inferenceTime;
+    }
+
+    public Bitmap inputImage() {
+        return inputImage;
+    }
+
+    public Bitmap outputImage() {
+        return outputImage;
+    }
+
+    public String outputResult() {
+        return outputResult;
+    }
+
+    public float preprocessTime() {
+        return preprocessTime;
+    }
+
+    public float postprocessTime() {
+        return postprocessTime;
+    }
+
+
+    public void setInputImage(Bitmap image) {
+        if (image == null) {
+            return;
+        }
+        this.inputImage = image.copy(Bitmap.Config.ARGB_8888, true);
+    }
+
+    private ArrayList<OcrResultModel> postprocess(ArrayList<OcrResultModel> results) {
+        for (OcrResultModel r : results) {
+            StringBuffer word = new StringBuffer();
+            for (int index : r.getWordIndex()) {
+                if (index >= 0 && index < wordLabels.size()) {
+                    word.append(wordLabels.get(index));
+                } else {
+                    Log.e(TAG, "Word index is not in label list:" + index);
+                    word.append("×");
+                }
+            }
+            r.setLabel(word.toString());
+        }
+        return results;
+    }
+
+    private void drawResults(ArrayList<OcrResultModel> results) {
+        StringBuffer outputResultSb = new StringBuffer("");
+        for (int i = 0; i < results.size(); i++) {
+            OcrResultModel result = results.get(i);
+            StringBuilder sb = new StringBuilder("");
+            sb.append(result.getLabel());
+            sb.append(" ").append(result.getConfidence());
+            sb.append("; Points: ");
+            for (Point p : result.getPoints()) {
+                sb.append("(").append(p.x).append(",").append(p.y).append(") ");
+            }
+            Log.i(TAG, sb.toString()); // show LOG in Logcat panel
+            outputResultSb.append(i + 1).append(": ").append(result.getLabel()).append("\n");
+        }
+        outputResult = outputResultSb.toString();
+        outputImage = inputImage;
+        Canvas canvas = new Canvas(outputImage);
+        Paint paintFillAlpha = new Paint();
+        paintFillAlpha.setStyle(Paint.Style.FILL);
+        paintFillAlpha.setColor(Color.parseColor("#3B85F5"));
+        paintFillAlpha.setAlpha(50);
+
+        Paint paint = new Paint();
+        paint.setColor(Color.parseColor("#3B85F5"));
+        paint.setStrokeWidth(5);
+        paint.setStyle(Paint.Style.STROKE);
+
+        for (OcrResultModel result : results) {
+            Path path = new Path();
+            List<Point> points = result.getPoints();
+            path.moveTo(points.get(0).x, points.get(0).y);
+            for (int i = points.size() - 1; i >= 0; i--) {
+                Point p = points.get(i);
+                path.lineTo(p.x, p.y);
+            }
+            canvas.drawPath(path, paint);
+            canvas.drawPath(path, paintFillAlpha);
+        }
+    }
+
+}

deploy/android_demo/app/src/main/res/menu/menu_action_options.xml

+<menu xmlns:android="http://schemas.android.com/apk/res/android"
+      xmlns:app="http://schemas.android.com/apk/res-auto">
+
+    <group>
+        <item
+                android:id="@+id/settings"
+                android:title="Settings..."
+                app:showAsAction="withText"/>
+    </group>
+</menu>

deploy/android_demo/app/src/main/res/mipmap-anydpi-v26/ic_launcher.xml

+<?xml version="1.0" encoding="utf-8"?>
+<adaptive-icon xmlns:android="http://schemas.android.com/apk/res/android">
+    <background android:drawable="@drawable/ic_launcher_background" />
+    <foreground android:drawable="@drawable/ic_launcher_foreground" />
+</adaptive-icon>
\ No newline at end of file

deploy/android_demo/app/src/main/res/mipmap-anydpi-v26/ic_launcher_round.xml

+<?xml version="1.0" encoding="utf-8"?>
+<adaptive-icon xmlns:android="http://schemas.android.com/apk/res/android">
+    <background android:drawable="@drawable/ic_launcher_background" />
+    <foreground android:drawable="@drawable/ic_launcher_foreground" />
+</adaptive-icon>
\ No newline at end of file