diff --git a/deploy/cpp_infer/include/preprocess_op.h b/deploy/cpp_infer/include/preprocess_op.h index bda7d1230fdc11c3b1a7d99130df755d26895de8..a19bac9500fb791d13ee8ce995af83c26112ff20 100644 --- a/deploy/cpp_infer/include/preprocess_op.h +++ b/deploy/cpp_infer/include/preprocess_op.h @@ -62,4 +62,4 @@ public: const std::vector &rec_image_shape = {3, 48, 192}); }; -} // namespace PaddleOCR \ No newline at end of file +} // namespace PaddleOCR diff --git a/deploy/cpp_infer/src/preprocess_op.cpp b/deploy/cpp_infer/src/preprocess_op.cpp index 1d3e31a14888969b0b56398bdcb9925d03e91cfe..e2f87e5a5adbba874d16199c93eb499f5fffc136 100644 --- a/deploy/cpp_infer/src/preprocess_op.cpp +++ b/deploy/cpp_infer/src/preprocess_op.cpp @@ -81,7 +81,7 @@ void ResizeImgType0::Run(const cv::Mat &img, cv::Mat &resize_img, else if (resize_h / 32 < 1 + 1e-5) resize_h = 32; else - resize_h = (resize_h / 32 - 1) * 32; + resize_h = resize_h / 32 * 32; if (resize_w % 32 == 0) resize_w = resize_w; @@ -96,13 +96,17 @@ void ResizeImgType0::Run(const cv::Mat &img, cv::Mat &resize_img, ratio_w = float(resize_w) / float(w); } -void CrnnResizeImg::Run(const cv::Mat &img, cv::Mat &resize_img, float wh_ratio, +void CrnnResizeImg::Run(const cv::Mat &img, cv::Mat &pad_resize_img, + float max_wh_ratio, const std::vector &rec_image_shape) { int imgC, imgH, imgW; imgC = rec_image_shape[0]; imgH = rec_image_shape[1]; imgW = rec_image_shape[2]; + float wh_ratio = 1.0 * imgW / imgH; + wh_ratio = std::max(max_wh_ratio, wh_ratio); + imgW = int(32 * wh_ratio); float ratio = float(img.cols) / float(img.rows); @@ -112,8 +116,12 @@ void CrnnResizeImg::Run(const cv::Mat &img, cv::Mat &resize_img, float wh_ratio, 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_LINEAR); + cv::copyMakeBorder(resize_img, pad_resize_img, 0, 0, 0, + int(imgW - resize_img.cols), cv::BORDER_CONSTANT, + {127, 127, 127}); } void ClsResizeImg::Run(const cv::Mat &img, cv::Mat &resize_img, diff --git a/deploy/lite/readme.md b/deploy/lite/readme.md index e678b2f6164dbfb82c690b55b43f9f66055a24ef..33d98b6cc419470658ba8c34c6ab20b317fd0bfc 100644 --- a/deploy/lite/readme.md +++ b/deploy/lite/readme.md @@ -24,15 +24,7 @@ Paddle Lite是飞桨轻量化推理引擎，为手机、IOT端提供高效推理 ### 1.2 准备预测库 预测库有两种获取方式： -- 1. 直接下载，预测库下载链接如下： - |平台|预测库下载链接| - |-|-| - |Android|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.6.3/inference_lite_lib.android.armv7.gcc.c++_shared.with_extra.with_cv.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.6.3/inference_lite_lib.android.armv8.gcc.c++_shared.with_extra.with_cv.tar.gz)| - |IOS|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.6.3/inference_lite_lib.ios.armv7.with_cv.with_extra.with_log.tiny_publish.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.6.3/inference_lite_lib.ios.armv8.with_cv.with_extra.with_log.tiny_publish.tar.gz)| - - 注：1. 上述预测库为PaddleLite 2.6.3分支编译得到，有关PaddleLite 2.6.3 详细信息可参考[链接](https://github.com/PaddlePaddle/Paddle-Lite/releases/tag/v2.6.3)。 - -- 2. [推荐]编译Paddle-Lite得到预测库，Paddle-Lite的编译方式如下： +- 1. [推荐]编译Paddle-Lite得到预测库，Paddle-Lite的编译方式如下： ``` git clone https://github.com/PaddlePaddle/Paddle-Lite.git cd Paddle-Lite @@ -45,6 +37,9 @@ git checkout release/v2.7 更多编译命令 介绍请参考[链接](https://paddle-lite.readthedocs.io/zh/latest/user_guides/Compile/Android.html#id2)。 +- 2. 直接下载预测库，下载[链接](https://github.com/PaddlePaddle/Paddle-Lite/releases/tag/v2.7.1)。 + + 直接下载预测库并解压后，可以得到`inference_lite_lib.android.armv8/`文件夹，通过编译Paddle-Lite得到的预测库位于 `Paddle-Lite/build.lite.android.armv8.gcc/inference_lite_lib.android.armv8/`文件夹下。 预测库的文件目录如下： @@ -88,11 +83,11 @@ Paddle-Lite 提供了多种策略来自动优化原始的模型，其中包括 |-|-|-|-|-|-|-| |V1.1|超轻量中文OCR 移动端模型|8.1M|[下载地址](https://paddleocr.bj.bcebos.com/20-09-22/mobile/lite/ch_ppocr_mobile_v1.1_det_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/20-09-22/mobile/lite/ch_ppocr_mobile_v1.1_cls_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/20-09-22/mobile/lite/ch_ppocr_mobile_v1.1_rec_opt.nb)|v2.7| |【slim】V1.1|超轻量中文OCR 移动端模型|3.5M|[下载地址](https://paddleocr.bj.bcebos.com/20-09-22/mobile/lite/ch_ppocr_mobile_v1.1_det_prune_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/20-09-22/mobile/lite/ch_ppocr_mobile_v1.1_cls_quant_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/20-09-22/mobile/lite/ch_ppocr_mobile_v1.1_rec_quant_opt.nb)|v2.7| -|V1.0|轻量级中文OCR 移动端模型|8.6M|[下载地址](https://paddleocr.bj.bcebos.com/20-09-22/mobile/lite/ch_ppocr_mobile_v1.0_det_opt.nb)|---|[下载地址](https://paddleocr.bj.bcebos.com/20-09-22/mobile/lite/ch_ppocr_mobile_v1.0_rec_opt.nb)|v2.7| + 注意：V1.1 3.0M 轻量模型是使用PaddleSlim优化后的，需要配合Paddle-Lite最新预测库使用。 -如果直接使用上述表格中的模型进行部署，可略过下述步骤，直接阅读 [2.2节](#2.2与手机联调)。 +如果直接使用上述表格中的模型进行部署没有问题，可略过下述步骤，直接阅读 [2.2节](#2.2与手机联调)。 如果要部署的模型不在上述表格中，则需要按照如下步骤获得优化后的模型。 @@ -184,7 +179,7 @@ wget https://paddleocr.bj.bcebos.com/ch_models/ch_rec_mv3_crnn_infer.tar && tar ``` git clone https://github.com/PaddlePaddle/PaddleOCR.git cd PaddleOCR/deploy/lite/ - # 运行prepare.sh，准备预测库文件、测试图像和使用的字典文件，并放置在预测库中的demo/cxx/ocr文件夹下 + # 按照如下命令运行prepare.sh，将预测库文件、测试图像和使用的字典文件到预测库中的demo/cxx/ocr文件夹下 sh prepare.sh /{lite prediction library path}/inference_lite_lib.android.armv8 # 进入OCR demo的工作目录 @@ -255,6 +250,7 @@ use_direction_classify 0 # 是否使用方向分类器，0表示不使用，1 adb shell cd /data/local/tmp/debug export LD_LIBRARY_PATH=${PWD}:$LD_LIBRARY_PATH + # ./ocr_db_crnn 检测模型文件 方向分类器模型文件 识别模型文件 测试图像路径 字典文件路径 ./ocr_db_crnn ch_ppocr_mobile_v1.1_det_prune_opt.nb ch_ppocr_mobile_v1.1_rec_quant_opt.nb ch_ppocr_mobile_cls_quant_opt.nb ./11.jpg ppocr_keys_v1.txt ``` diff --git a/deploy/lite/readme_en.md b/deploy/lite/readme_en.md index 0780525420c41f3609a17f50de4668a9b1f2568e..9df0ee375b35bf6a44645e0c7e16d3254c80aa8c 100644 --- a/deploy/lite/readme_en.md +++ b/deploy/lite/readme_en.md @@ -22,15 +22,7 @@ deployment solutions for end-side deployment issues. ## 3. Prepare prebuild library for android and ios -### 3.1 Download prebuild library -|Platform|Prebuild library Download Link| -|-|-| -|Android|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.6.3/inference_lite_lib.android.armv7.gcc.c++_shared.with_extra.with_cv.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.6.3/inference_lite_lib.android.armv8.gcc.c++_shared.with_extra.with_cv.tar.gz)| -|IOS|[arm7](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.6.3/inference_lite_lib.ios.armv7.with_cv.with_extra.with_log.tiny_publish.tar.gz) / [arm8](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.6.3/inference_lite_lib.ios.armv8.with_cv.with_extra.with_log.tiny_publish.tar.gz)| - -note: The above pre-build inference library is compiled from the PaddleLite `release/v2.7` branch. For more information about PaddleLite 2.6.3, please refer to [link](https://github.com/PaddlePaddle/Paddle-Lite/releases/tag/v2.6.3). - -### 3.2 Compile prebuild library (Recommended) +### 3.1 Compile prebuild library (Recommended) ``` git clone https://github.com/PaddlePaddle/Paddle-Lite.git cd Paddle-Lite @@ -66,6 +58,11 @@ inference_lite_lib.android.armv8/ | `-- java ``` +### 3.2 Download prebuild library + +PaddleLite also provides a compiled [prediction library](https://github.com/PaddlePaddle/Paddle-Lite/releases/tag/v2.7.1), developers can try on their own. + + ## 4. Inference Model Optimization @@ -80,7 +77,6 @@ You can directly download the optimized model. | - | - | - | - | - | - | - | | V1.1 | extra-lightweight chinese OCR optimized model | 8.1M | [Download](https://paddleocr.bj.bcebos.com/20-09-22/mobile/lite/ch_ppocr_mobile_v1.1_det_opt.nb) | [Download](https://paddleocr.bj.bcebos.com/20-09-22/mobile/lite/ch_ppocr_mobile_v1.1_cls_opt.nb) | [Download](https://paddleocr.bj.bcebos.com/20-09-22/mobile/lite/ch_ppocr_mobile_v1.1_rec_opt.nb) | develop | | [slim] V1.1 | extra-lightweight chinese OCR optimized model | 3.5M | [Download](https://paddleocr.bj.bcebos.com/20-09-22/mobile/lite/ch_ppocr_mobile_v1.1_det_prune_opt.nb) | [Download](https://paddleocr.bj.bcebos.com/20-09-22/mobile/lite/ch_ppocr_mobile_v1.1_cls_quant_opt.nb) | [Download](https://paddleocr.bj.bcebos.com/20-09-22/mobile/lite/ch_ppocr_mobile_v1.1_rec_quant_opt.nb) | develop | -| V1.0 | lightweight Chinese OCR optimized model | 8.6M | [Download](https://paddleocr.bj.bcebos.com/20-09-22/mobile/lite/ch_ppocr_mobile_v1.0_det_opt.nb) | - | [Download](https://paddleocr.bj.bcebos.com/20-09-22/mobile/lite/ch_ppocr_mobile_v1.0_rec_opt.nb) | develop | If the model to be deployed is not in the above table, you need to follow the steps below to obtain the optimized model. diff --git a/doc/doc_en/algorithm_overview_en.md b/doc/doc_en/algorithm_overview_en.md index 2e21fd621971e062384a9323e79a8cf4498d7495..e23184969ae09f80925be1735652cf92d5a2992b 100644 --- a/doc/doc_en/algorithm_overview_en.md +++ b/doc/doc_en/algorithm_overview_en.md @@ -33,7 +33,7 @@ On Total-Text dataset, the text detection result is as follows: **Note：** Additional data, like icdar2013, icdar2017, COCO-Text, ArT, was added to the model training of SAST. Download English public dataset in organized format used by PaddleOCR from [Baidu Drive](https://pan.baidu.com/s/12cPnZcVuV1zn5DOd4mqjVw) (download code: 2bpi). -For the training guide and use of PaddleOCR text detection algorithms, please refer to the document [Text detection model training/evaluation/prediction](./doc/doc_en/detection_en.md) +For the training guide and use of PaddleOCR text detection algorithms, please refer to the document [Text detection model training/evaluation/prediction](./detection_en.md) ### 2. Text Recognition Algorithm @@ -63,4 +63,4 @@ Refer to [DTRB](https://arxiv.org/abs/1904.01906), the training and evaluation r The average accuracy of the two-stage training in the original paper is 89.74%, and that of one stage training in paddleocr is 88.33%. Both pre-trained weights can be downloaded [here](https://paddleocr.bj.bcebos.com/SRN/rec_r50fpn_vd_none_srn.tar). -Please refer to the document for training guide and use of PaddleOCR text recognition algorithms [Text recognition model training/evaluation/prediction](./doc/doc_en/recognition_en.md) +Please refer to the document for training guide and use of PaddleOCR text recognition algorithms [Text recognition model training/evaluation/prediction](./recognition_en.md) diff --git a/ppocr/modeling/backbones/det_mobilenet_v3.py b/ppocr/modeling/backbones/det_mobilenet_v3.py index 508f2bbf2d73216fc35a2d906f94731707a25736..0b1f5dc0cb107d9f6f646c8449d4886d4f63c4f6 100755 --- a/ppocr/modeling/backbones/det_mobilenet_v3.py +++ b/ppocr/modeling/backbones/det_mobilenet_v3.py @@ -32,6 +32,7 @@ class MobileNetV3(): """ self.scale = params['scale'] model_name = params['model_name'] + self.model_name = model_name self.inplanes = 16 if model_name == "large": self.cfg = [ @@ -80,7 +81,7 @@ class MobileNetV3(): "supported scale are {} but input scale is {}".format(supported_scale, self.scale) self.disable_se = params.get('disable_se', False) - + def __call__(self, input): scale = self.scale inplanes = self.inplanes @@ -102,7 +103,8 @@ class MobileNetV3(): inplanes = self.make_divisible(inplanes * scale) outs = [] for layer_cfg in cfg: - if layer_cfg[5] == 2 and i > 2: + start_idx = 2 if self.model_name == 'large' else 0 + if layer_cfg[5] == 2 and i > start_idx: outs.append(conv) conv = self.residual_unit( input=conv, diff --git a/tools/infer/predict_det.py b/tools/infer/predict_det.py index 21f9bf90b9546dfb86266a07683d4af1200d53fa..942e23643887568678a8f5cb32c486e18866c201 100755 --- a/tools/infer/predict_det.py +++ b/tools/infer/predict_det.py @@ -200,4 +200,4 @@ if __name__ == "__main__": logger.info("The visualized img saved in {}".format( os.path.join(draw_img_save, "det_res_%s" % img_name_pure))) if count > 1: - logger.info("Avg Time:", total_time / (count - 1)) + logger.info("Avg Time: {}".format(total_time / (count - 1)))