提交 399a6358 编写于 作者: A andyjpaddle

Merge branch 'dygraph' of https://github.com/PaddlePaddle/PaddleOCR into dygraph

...@@ -36,8 +36,8 @@ op: ...@@ -36,8 +36,8 @@ op:
#det模型路径 #det模型路径
model_config: ./ppocr_det_v3_serving model_config: ./ppocr_det_v3_serving
#Fetch结果列表,以client_config中fetch_var的alias_name为准 #Fetch结果列表,以client_config中fetch_var的alias_name为准,不设置默认取全部输出变量
fetch_list: ["sigmoid_0.tmp_0"] #fetch_list: ["sigmoid_0.tmp_0"]
#计算硬件ID,当devices为""或不写时为CPU预测;当devices为"0", "0,1,2"时为GPU预测,表示使用的GPU卡 #计算硬件ID,当devices为""或不写时为CPU预测;当devices为"0", "0,1,2"时为GPU预测,表示使用的GPU卡
devices: "0" devices: "0"
...@@ -62,8 +62,8 @@ op: ...@@ -62,8 +62,8 @@ op:
#rec模型路径 #rec模型路径
model_config: ./ppocr_rec_v3_serving model_config: ./ppocr_rec_v3_serving
#Fetch结果列表,以client_config中fetch_var的alias_name为准 #Fetch结果列表,以client_config中fetch_var的alias_name为准, 不设置默认取全部输出变量
fetch_list: ["softmax_5.tmp_0"] #fetch_list:
#计算硬件ID,当devices为""或不写时为CPU预测;当devices为"0", "0,1,2"时为GPU预测,表示使用的GPU卡 #计算硬件ID,当devices为""或不写时为CPU预测;当devices为"0", "0,1,2"时为GPU预测,表示使用的GPU卡
devices: "0" devices: "0"
......
...@@ -393,7 +393,7 @@ class OCRReader(object): ...@@ -393,7 +393,7 @@ class OCRReader(object):
return norm_img_batch[0] return norm_img_batch[0]
def postprocess(self, outputs, with_score=False): def postprocess(self, outputs, with_score=False):
preds = outputs["softmax_5.tmp_0"] preds = list(outputs.values())[0]
try: try:
preds = preds.numpy() preds = preds.numpy()
except: except:
...@@ -404,8 +404,11 @@ class OCRReader(object): ...@@ -404,8 +404,11 @@ class OCRReader(object):
preds_idx, preds_prob, is_remove_duplicate=True) preds_idx, preds_prob, is_remove_duplicate=True)
return text return text
from argparse import ArgumentParser,RawDescriptionHelpFormatter
from argparse import ArgumentParser, RawDescriptionHelpFormatter
import yaml import yaml
class ArgsParser(ArgumentParser): class ArgsParser(ArgumentParser):
def __init__(self): def __init__(self):
super(ArgsParser, self).__init__( super(ArgsParser, self).__init__(
...@@ -441,7 +444,7 @@ class ArgsParser(ArgumentParser): ...@@ -441,7 +444,7 @@ class ArgsParser(ArgumentParser):
s = s.strip() s = s.strip()
k, v = s.split('=') k, v = s.split('=')
v = self._parse_helper(v) v = self._parse_helper(v)
print(k,v, type(v)) print(k, v, type(v))
cur = config cur = config
parent = cur parent = cur
for kk in k.split("."): for kk in k.split("."):
......
...@@ -56,7 +56,7 @@ class DetOp(Op): ...@@ -56,7 +56,7 @@ class DetOp(Op):
return {"x": det_img[np.newaxis, :].copy()}, False, None, "" return {"x": det_img[np.newaxis, :].copy()}, False, None, ""
def postprocess(self, input_dicts, fetch_dict, data_id, log_id): def postprocess(self, input_dicts, fetch_dict, data_id, log_id):
det_out = fetch_dict["sigmoid_0.tmp_0"] det_out = list(fetch_dict.values())[0]
ratio_list = [ ratio_list = [
float(self.new_h) / self.ori_h, float(self.new_w) / self.ori_w float(self.new_h) / self.ori_h, float(self.new_w) / self.ori_w
] ]
......
...@@ -55,7 +55,7 @@ class DetOp(Op): ...@@ -55,7 +55,7 @@ class DetOp(Op):
return {"x": det_img[np.newaxis, :].copy()}, False, None, "" return {"x": det_img[np.newaxis, :].copy()}, False, None, ""
def postprocess(self, input_dicts, fetch_dict, data_id, log_id): def postprocess(self, input_dicts, fetch_dict, data_id, log_id):
det_out = fetch_dict["sigmoid_0.tmp_0"] det_out = list(fetch_dict.values())[0]
ratio_list = [ ratio_list = [
float(self.new_h) / self.ori_h, float(self.new_w) / self.ori_w float(self.new_h) / self.ori_h, float(self.new_w) / self.ori_w
] ]
......
...@@ -8,123 +8,214 @@ ...@@ -8,123 +8,214 @@
- [4. 端到端评估](#4) - [4. 端到端评估](#4)
<a name="1"></a> <a name="1"></a>
## 1. 简介 ## 1. 简介
PP-OCRv3在PP-OCRv2的基础上进一步升级。检测模型仍然基于DB算法,优化策略采用了带残差注意力机制的FPN结构RSEFPN、增大感受野的PAN结构LKPAN、基于DML训练的更优的教师模型;识别模型将base模型从CRNN替换成了IJCAI 2022论文[SVTR](),并采用SVTR轻量化、带指导训练CTC、数据增广策略RecConAug、自监督训练的更好的预训练模型、无标签数据的使用进行模型加速和效果提升。更多细节请参考PP-OCRv3[技术报告](./PP-OCRv3_introduction.md)。 PP-OCRv3在PP-OCRv2的基础上进一步升级。整体的框架图保持了与PP-OCRv2相同的pipeline,针对检测模型和识别模型进行了优化。其中,检测模型仍基于DB模型优化,而识别模型不再采用CRNN,换成了会议IJCAI 2022中的最新方法[SVTR](https://arxiv.org/abs/2205.00159),PP-OCRv3系统框图如下所示(粉色框中为PP-OCRv3新增策略):
PP-OCRv3系统pipeline如下:
<div align="center"> <div align="center">
<img src="../ppocrv3_framework.png" width="800"> <img src="../ppocrv3_framework.png" width="800">
</div> </div>
从算法改进思路上看,分别针对检测和识别模型,进行了共八个方面的改进:
- 检测模型优化:
- LK-PAN:增大感受野的PAN模块;
- DML:教师模型互学习策略;
- RSE-FPN:带残差注意力机制的FPN模块;
- 识别模型优化:
- SVTR_LCNet:轻量级文本识别网络;
- GTC:Attention指导CTC训练策略;
- TextConAug:丰富图像上下文信息的数据增广策略;
- TextRotNet:自监督的预训练模型;
- UIM:无标签数据挖掘方案。
从效果上看,速度可比情况下,多种场景精度均有大幅提升:
- 中文场景,相对于PP-OCRv2中文模型提升超5%;
- 英文数字场景,相比于PP-OCRv2英文模型提升11%;
- 多语言场景,优化80+语种识别效果,平均准确率提升超5%。
<a name="2"></a> <a name="2"></a>
## 2. 检测优化 ## 2. 检测优化
PP-OCRv3采用PP-OCRv2的[CML](https://arxiv.org/pdf/2109.03144.pdf)蒸馏策略,在蒸馏的student模型、teacher模型精度提升,CML蒸馏策略上分别做了优化 PP-OCRv3检测模型整体训练方案仍采用PP-OCRv2的[CML](https://arxiv.org/pdf/2109.03144.pdf)蒸馏策略,CML蒸馏包含一个教师模型和两个学生模型,在训练过程中,教师模型不参与训练,学生模型受到来自标签和教师模型的监督,同时两个学生模型互相学习。PP-OCRv3分别针对教师模型、学生模型进一步优化。其中,在对教师模型优化时,采用了增大感受野的PAN模块LK-PAN和DML蒸馏策略;在对学生模型优化时,采用了带残差注意力机制的FPN模块RSE-FPN
- 在蒸馏student模型精度提升方面,提出了基于残差结构的通道注意力模块RSEFPN(Residual Squeeze-and-Excitation FPN),用于提升student模型精度和召回。 PP-OCRv3 CML蒸馏训练框架图如下:
RSEFPN的网络结构如下图所示,RSEFPN在PP-OCRv2的FPN基础上,将FPN中的卷积层更换为了通道注意力结构的RSEConv层。
<div align="center"> <div align="center">
<img src=".././ppocr_v3/RSEFPN.png" width="800"> <img src=".././ppocr_v3/ppocrv3_det_cml.png" width="800">
</div> </div>
RSEFPN将PP-OCR检测模型的精度hmean从81.3%提升到84.5%。模型大小从3M变为3.6M。 消融实验如下:
*注:PP-OCRv2的FPN通道数仅为96和24,如果直接用SE模块代替FPN的卷积会导致精度下降,RSEConv引入残差结构可以防止训练中包含重要特征的通道被抑制。* |序号|策略|模型大小|hmean|速度(cpu + mkldnn)|
|-|-|-|-|-|
|baseline teacher|PP-OCR server|49M|83.2%|171ms|
|teacher1|DB-R50-LK-PAN|124M|85.0%|396ms|
|teacher2|DB-R50-LK-PAN-DML|124M|86.0%|396ms|
|baseline student|PP-OCRv2|3M|83.2%|117ms|
|student0|DB-MV3-RSE-FPN|3.6M|84.5%|124ms|
|student1|DB-MV3-CML(teacher2)|3M|84.3%|117ms|
|student2|DB-MV3-RSE-FPN-CML(teacher2)|3.6M|85.4%|124ms|
- 在蒸馏的teacher模型精度提升方面,提出了LKPAN结构替换PP-OCRv2的FPN结构,并且使用ResNet50作为Backbone,更大的模型带来更多的精度提升。另外,对teacher模型使用[DML](https://arxiv.org/abs/1706.00384)蒸馏策略进一步提升teacher模型的精度。最终teacher的模型指标相比ppocr_server_v2.0从83.2%提升到了86.0% 测试环境: Intel Gold 6148 CPU,预测时开启MKLDNN加速
*注:[PP-OCRv2的FPN结构](https://github.com/PaddlePaddle/PaddleOCR/blob/77acb3bfe51c8a46c684527f73cd218cefedb4a3/ppocr/modeling/necks/db_fpn.py#L107)对DB算法FPN结构做了轻量级设计* **(1)增大感受野的PAN模块LK-PAN(Large Kernel PAN)**
LKPAN的网络结构如下图所示: LK-PAN(Large Kernel PAN)是一个具有更大感受野的轻量级[PAN](https://arxiv.org/pdf/1803.01534.pdf)结构。在LK-PAN的path augmentation中,使用卷积核为`9*9`的卷积;更大的卷积核意味着更大的感受野,更容易检测大字体的文字以及极端长宽比的文字。LK-PAN将PP-OCR server检测模型的hmean从83.2%提升到85.0%。
<div align="center"> <div align="center">
<img src="../ppocr_v3/LKPAN.png" width="800"> <img src="../ppocr_v3/LKPAN.png" width="1000">
</div> </div>
LKPAN(Large Kernel PAN)是一个具有更大感受野的轻量级[PAN](https://arxiv.org/pdf/1803.01534.pdf)结构。在LKPAN的path augmentation中,使用kernel size为`9*9`的卷积;更大的kernel size意味着更大的感受野,更容易检测大字体的文字以及极端长宽比的文字。LKPAN将PP-OCR检测模型的精度hmean从81.3%提升到84.9%。 **(2)DML(Deep Mutual Learning)蒸馏进一步提升teacher模型精度。**
*注:LKPAN相比RSEFPN有更多的精度提升,但是考虑到模型大小和预测速度等因素,在student模型中使用RSEFPN。* [DML](https://arxiv.org/abs/1706.00384) 互学习蒸馏方法,通过两个结构相同的模型互相学习,相比于传统的教师模型监督学生模型的蒸馏方法,DML 摆脱了对大的教师模型的依赖,蒸馏训练的流程更加简单。在PP-OCRv3的检测模型训练中,使用DML蒸馏策略进一步提升教师模型的精度,并使用ResNet50作为Backbone。DML策略将教师模型的Hmean从85%进一步提升至86%。
采用上述策略,PP-OCRv3相比PP-OCRv2,hmean指标从83.3%提升到85.4%;预测速度从平均117ms/image变为124ms/image。 教师模型DML训练流程图如下:
3. PP-OCRv3检测模型消融实验 <div align="center">
<img src="../ppocr_v3/teacher_dml.png" width="800">
</div>
|序号|策略|模型大小|hmean|Intel Gold 6148CPU+mkldnn预测耗时| **(3)带残差注意力机制的FPN模块RSE-FPN(Residual SE-FPN)。**
|-|-|-|-|-|
|0|PP-OCR|3M|81.3%|117ms| 残差结构的通道注意力模块RSE-FPN结构如下图所示,RSE-FPN在PP-OCRv2的FPN基础上,将FPN中的卷积层更换为通道注意力结构的RSEConv层。考虑到PP-OCRv2的FPN通道数仅为96和24,如果直接用SEblock代替FPN中卷积会导致某些通道的特征被抑制,进而导致精度下降,RSEConv引入残差结构防止训练中包含重要特征的通道被抑制。直接添加RSE-FPN模块,可将PP-OCR检测模型的精度Hmean从81.3%提升到84.5%。在学生模型中加入RSE-FPN后进行CML蒸馏,比不加时,Hmean指标从83.2提升到84.3%。
|1|PP-OCRV2|3M|83.3%|117ms|
|2|0 + RESFPN|3.6M|84.5%|124ms|
|3|0 + LKPAN|4.6M|84.9%|156ms|
|4|ppocr_server_v2.0 |124M|83.2%||171ms|
|5|teacher + DML + LKPAN|124M|86.0%|396ms|
|6|0 + 2 + 5 + CML|3.6M|85.4%|124ms|
<div align="center">
<img src=".././ppocr_v3/RSEFPN.png" width="1000">
</div>
<a name="3"></a> <a name="3"></a>
## 3. 识别优化 ## 3. 识别优化
[SVTR](https://arxiv.org/abs/2205.00159) 证明了强大的单视觉模型(无需序列模型)即可高效准确完成文本识别任务,在中英文数据上均有优秀的表现。经过实验验证,SVTR_Tiny在自建的 [中文数据集上](https://arxiv.org/abs/2109.03144) ,识别精度可以提升10.7%,网络结构如下所示: PP-OCRv3识别模型从网络结构、训练策略、数据增广等多个方面进行了优化,PP-OCRv3系统流程图如下所示:
<img src="../ppocr_v3/svtr_tiny.jpg" width=800> <div align="center">
<img src="../ppocr_v3/v3_rec_pipeline.png" width=800>
</div>
由于 MKLDNN 加速库支持的模型结构有限,SVTR 在CPU+MKLDNN上相比PP-OCRv2慢了10倍 上图中,蓝色方块中列举了PP-OCRv3识别模型的6个主要模块。首先在模块①,将base模型从CRNN替换为精度更高的单一视觉模型[SVTR](https://arxiv.org/abs/2205.00159),并进行一系列的结构优化进行加速,得到全新的轻量级文本识别网络SVTR_LCNet(如图中红色虚线框所示);在模块②,借鉴[GTC](https://arxiv.org/pdf/2002.01276.pdf)策略,引入Attention指导CTC训练,进一步提升模型精度;在模块③,使用基于上下文信息的数据增广策略TextConAug,丰富训练数据上下文信息,提升训练数据多样性;在模块④,使用TextRotNet训练自监督的预训练模型,充分利用无标注识别数据的信息;模块⑤基于PP-OCRv2中提出的UDML蒸馏策略进行蒸馏学习,除计算2个模型的CTC分支的DMLLoss外,也计算2个模型的Attention分支之间的DMLLoss,从而得到更优模型;在模块⑥中,基于UIM无标注数据挖掘方法,使用效果好但速度相对较慢的SVTR_tiny模型进行无标签数据挖掘,为模型训练增加更多真实数据
PP-OCRv3 期望在提升模型精度的同时,不带来额外的推理耗时。通过分析发现,SVTR_Tiny结构的主要耗时模块为Mixing Block,因此我们对 SVTR_Tiny 的结构进行了一系列优化(详细速度数据请参考下方消融实验表格):
1. 将SVTR网络前半部分替换为PP-LCNet的前三个stage,保留4个 Global Mixing Block ,精度为76%,加速69%,网络结构如下所示: 基于上述策略,PP-OCRv3识别模型相比PP-OCRv2,在速度可比的情况下,精度进一步提升4.6%。 具体消融实验如下所示:
<img src="../ppocr_v3/svtr_g4.png" width=800>
2. 将4个 Global Attenntion Block 减小到2个,精度为72.9%,加速69%,网络结构如下所示:
<img src="../ppocr_v3/svtr_g2.png" width=800>
3. 实验发现 Global Attention 的预测速度与输入其特征的shape有关,因此后移Global Mixing Block的位置到池化层之后,精度下降为71.9%,速度超越 CNN-base 的PP-OCRv2 22%,网络结构如下所示:
<img src="../ppocr_v3/ppocr_v3.png" width=800>
为了提升模型精度同时不引入额外推理成本,PP-OCRv3参考GTC策略,使用Attention监督CTC训练,预测时完全去除Attention模块,在推理阶段不增加任何耗时, 精度提升3.8%,训练流程如下所示: | ID | 策略 | 模型大小 | 精度 | 预测耗时(CPU + MKLDNN)|
<img src="../ppocr_v3/GTC.png" width=800> |-----|-----|--------|----| --- |
| 01 | PP-OCRv2 | 8M | 74.8% | 8.54ms |
| 02 | SVTR_Tiny | 21M | 80.1% | 97ms |
| 03 | SVTR_LCNet | 12M | 71.9% | 6.6ms |
| 04 | + GTC | 12M | 75.8% | 7.6ms |
| 05 | + TextConAug | 12M | 76.3% | 7.6ms |
| 06 | + TextRotNet | 12M | 76.9% | 7.6ms |
| 07 | + UDML | 12M | 78.4% | 7.6ms |
| 08 | + UIM | 12M | 79.4% | 7.6ms |
在训练策略方面,PP-OCRv3参考 [SSL](https://github.com/ku21fan/STR-Fewer-Labels) 设计了文本方向任务,训练了适用于文本识别的预训练模型,加速模型收敛过程,精度提升了0.6%; 使用UDML蒸馏策略,进一步提升精度1.5%,训练流程所示: 注: 测试速度时,实验01-03输入图片尺寸均为(3,32,320),04-08输入图片尺寸均为(3,48,320)。在实际预测时,图像为变长输入,速度会有所变化。
<img src="../ppocr_v3/SSL.png" width="300"> <img src="../ppocr_v3/UDML.png" width="500">
**(1)轻量级文本识别网络SVTR_LCNet。**
数据增强方面 PP-OCRv3将base模型从CRNN替换成了[SVTR](https://arxiv.org/abs/2205.00159),SVTR证明了强大的单视觉模型(无需序列模型)即可高效准确完成文本识别任务,在中英文数据上均有优秀的表现。经过实验验证,SVTR_Tiny 在自建的[中文数据集](https://arxiv.org/abs/2109.03144)上 ,识别精度可以提升至80.1%,SVTR_Tiny 网络结构如下所示
1. 基于 [ConCLR](https://www.cse.cuhk.edu.hk/~byu/papers/C139-AAAI2022-ConCLR.pdf) 中的ConAug方法,设计了 RecConAug 数据增强方法,增强数据多样性,精度提升0.5%,增强可视化效果如下所示: <div align="center">
<img src="../ppocr_v3/recconaug.png" width=800> <img src="../ppocr_v3/svtr_tiny.png" width=800>
</div>
2. 使用训练好的 SVTR_large 预测 120W 的 lsvt 无标注数据,取出其中得分大于0.95的数据,共得到81W识别数据加入到PP-OCRv3的训练数据中,精度提升1%。
总体来讲PP-OCRv3识别从网络结构、训练策略、数据增强三个方向做了进一步优化: 由于 MKLDNN 加速库支持的模型结构有限,SVTR 在 CPU+MKLDNN 上相比 PP-OCRv2 慢了10倍。PP-OCRv3 期望在提升模型精度的同时,不带来额外的推理耗时。通过分析发现,SVTR_Tiny 结构的主要耗时模块为 Mixing Block,因此我们对 SVTR_Tiny 的结构进行了一系列优化(详细速度数据请参考下方消融实验表格):
- 网络结构上:考虑[SVTR](https://arxiv.org/abs/2205.00159) 在中英文效果上的优越性,采用SVTR_Tiny作为base,选取Global Mixing Block和卷积组合提取特征,并将Global Mixing Block位置后移进行加速; 参考 [GTC](https://arxiv.org/pdf/2002.01276.pdf) 策略,使用注意力机制模块指导CTC训练,定位和识别字符,提升不规则文本的识别精度。
- 训练策略上:参考 [SSL](https://github.com/ku21fan/STR-Fewer-Labels) 设计了方向分类前序任务,获取更优预训练模型,加速模型收敛过程,提升精度; 使用UDML蒸馏策略、监督attention、ctc两个分支得到更优模型。
- 数据增强上:基于 [ConCLR](https://www.cse.cuhk.edu.hk/~byu/papers/C139-AAAI2022-ConCLR.pdf) 中的ConAug方法,改进得到 RecConAug 数据增广方法,支持随机结合任意多张图片,提升训练数据的上下文信息丰富度,增强模型鲁棒性;使用 SVTR_large 预测无标签数据,向训练集中补充81w高质量真实数据。
基于上述策略,PP-OCRv3识别模型相比PP-OCRv2,在速度可比的情况下,精度进一步提升4.5%。 具体消融实验如下所示: 1. 将 SVTR 网络前半部分替换为 PP-LCNet 的前三个stage,保留4个 Global Mixing Block ,精度为76%,加速69%,网络结构如下所示:
<div align="center">
<img src="../ppocr_v3/svtr_g4.png" width=800>
</div>
2. 将4个 Global Mixing Block 减小到2个,精度为72.9%,加速69%,网络结构如下所示:
<div align="center">
<img src="../ppocr_v3/svtr_g2.png" width=800>
</div>
3. 实验发现 Global Mixing Block 的预测速度与输入其特征的shape有关,因此后移 Global Mixing Block 的位置到池化层之后,精度下降为71.9%,速度超越基于CNN结构的PP-OCRv2-baseline 22%,网络结构如下所示:
<div align="center">
<img src="../ppocr_v3/LCNet_SVTR.png" width=800>
</div>
实验细节 具体消融实验如下所示
| id | 策略 | 模型大小 | 精度 | 速度(cpu + mkldnn)| | ID | 策略 | 模型大小 | 精度 | 速度(CPU + MKLDNN)|
|-----|-----|--------|----| --- | |-----|-----|--------|----| --- |
| 01 | PP-OCRv2 | 8M | 69.3% | 8.54ms | | 01 | PP-OCRv2-baseline | 8M | 69.3% | 8.54ms |
| 02 | SVTR_Tiny | 21M | 80.1% | 97ms | | 02 | SVTR_Tiny | 21M | 80.1% | 97ms |
| 03 | LCNet_SVTR_G4 | 9.2M | 76% | 30ms | | 03 | SVTR_LCNet(G4) | 9.2M | 76% | 30ms |
| 04 | LCNet_SVTR_G2 | 13M | 72.98% | 9.37ms | | 04 | SVTR_LCNet(G2) | 13M | 72.98% | 9.37ms |
| 05 | PP-OCRv3 | 12M | 71.9% | 6.6ms | | 05 | SVTR_LCNet | 12M | 71.9% | 6.6ms |
| 06 | + large input_shape | 12M | 73.98% | 7.6ms |
| 06 | + GTC | 12M | 75.8% | 7.6ms | 注: 测试速度时,输入图片尺寸均为(3,32,320); PP-OCRv2-baseline 代表没有借助蒸馏方法训练得到的模型
| 07 | + RecConAug | 12M | 76.3% | 7.6ms |
| 08 | + SSL pretrain | 12M | 76.9% | 7.6ms | **(2)采用Attention指导CTC训练。**
| 09 | + UDML | 12M | 78.4% | 7.6ms |
| 10 | + unlabeled data | 12M | 79.4% | 7.6ms | 为了提升模型精度同时不引入额外推理成本,PP-OCRv3 参考 GTC(Guided Training of CTC) 策略,使用 Attention 监督 CTC 训练,预测时完全去除 Attention 模块,在推理阶段不增加任何耗时, 精度提升3.8%,训练流程如下所示:
<div align="center">
注: 测试速度时,实验01-05输入图片尺寸均为(3,32,320),06-10输入图片尺寸均为(3,48,320) <img src="../ppocr_v3/GTC.png" width=800>
</div>
**(3)TextConAug数据增广策略。**
在论文[ConCLR](https://www.cse.cuhk.edu.hk/~byu/papers/C139-AAAI2022-ConCLR.pdf)中,作者提出ConAug数据增广,在一个batch内对2张不同的图像进行联结,组成新的图像并进行自监督对比学习。PP-OCRv3将此方法应用到有监督的学习任务中,设计了TextConAug数据增强方法,支持更多图像的联结,从而进一步丰富了图像的上下文信息。最终将识别模型精度进一步提升0.5%。TextConAug示意图如下所示:
<div align="center">
<img src="../ppocr_v3/recconaug.png" width=800>
</div>
**(4)TextRotNet自监督训练优化预训练模型。**
为了充分利用自然场景中的大量无标注文本数据,PP-OCRv3参考论文[STR-Fewer-Labels](https://github.com/ku21fan/STR-Fewer-Labels),设计TextRotNet自监督任务,对识别图像进行旋转并预测其旋转角度,同时结合中文场景文字识别任务的特点,在训练时适当调整图像的尺寸,添加文本识别数据增广,最终产出针对文本识别任务的PP-LCNet预训练模型,帮助识别模型精度进一步提升0.6%。TextRotNet训练流程如下图所示:
<div align="center">
<img src="../ppocr_v3/SSL.png" width="500">
</div>
**(5)UIM(Unlabeled Images Mining)无标注数据挖掘策略。**
为更直接利用自然场景中包含大量无标注数据,使用PP-OCRv2检测模型以及SVTR_tiny识别模型对百度开源的40W [LSVT弱标注数据集](https://ai.baidu.com/broad/introduction?dataset=lsvt)进行检测与识别,并筛选出识别得分大于0.95的文本,共81W文本行数据,将其补充到训练数据中,最终进一步提升模型精度1.0%。
<div align="center">
<img src="../ppocr_v3/UIM.png" width="500">
</div>
<a name="4"></a> <a name="4"></a>
## 4. 端到端评估 ## 4. 端到端评估
经过以上优化,最终PP-OCRv3在速度可比情况下,中文场景端到端Hmean指标相比于PP-OCRv2提升5%,效果大幅提升。具体指标如下表所示:
| Model | Hmean | Model Size (M) | Time Cost (CPU, ms) | Time Cost (T4 GPU, ms) |
|-----|-----|--------|----| --- |
| PP-OCR mobile | 50.3% | 8.1 | 356 | 116 |
| PP-OCR server | 57.0% | 155.1 | 1056 | 200 |
| PP-OCRv2 | 57.6% | 11.6 | 330 | 111 |
| PP-OCRv3 | 62.9% | 15.6 | 331 | 86.64 |
测试环境:CPU型号为Intel Gold 6148,CPU预测时开启MKLDNN加速。
除了更新中文模型,本次升级也同步优化了英文数字模型,端到端效果提升11%,如下表所示:
| Model | Recall | Precision | Hmean |
|-----|-----|--------|----|
| PP-OCR_en | 38.99% | 45.91% | 42.17% |
| PP-OCRv3_en | 50.95% | 55.53% | 53.14% |
同时,也对已支持的80余种语言识别模型进行了升级更新,在有评估集的四种语系识别准确率平均提升5%以上,如下表所示:
| Model | 拉丁语系 | 阿拉伯语系 | 日语 | 韩语 |
|-----|-----|--------|----| --- |
| PP-OCR_mul | 69.6% | 40.5% | 38.5% | 55.4% |
| PP-OCRv3_mul | 75.2%| 45.37% | 45.8% | 60.1% |
...@@ -41,7 +41,7 @@ PaddleOCR提供的可下载模型包括`推理模型`、`训练模型`、`预训 ...@@ -41,7 +41,7 @@ PaddleOCR提供的可下载模型包括`推理模型`、`训练模型`、`预训
|模型名称|模型简介|配置文件|推理模型大小|下载地址| |模型名称|模型简介|配置文件|推理模型大小|下载地址|
| --- | --- | --- | --- | --- | | --- | --- | --- | --- | --- |
|ch_PP-OCRv3_det_slim|【最新】slim量化+蒸馏版超轻量模型,支持中英文、多语种文本检测|[ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml)| 1.1M |[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_infer.tar) / [训练模型(coming soon)](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_distill_train.tar) / [slim模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_infer.nb)| |ch_PP-OCRv3_det_slim|【最新】slim量化+蒸馏版超轻量模型,支持中英文、多语种文本检测|[ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml)| 1.1M |[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_distill_train.tar) / [slim模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_infer.nb)|
|ch_PP-OCRv3_det| 【最新】原始超轻量模型,支持中英文、多语种文本检测 |[ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml)| 3.8M |[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_distill_train.tar)| |ch_PP-OCRv3_det| 【最新】原始超轻量模型,支持中英文、多语种文本检测 |[ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml)| 3.8M |[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_distill_train.tar)|
|ch_PP-OCRv2_det_slim| slim量化+蒸馏版超轻量模型,支持中英文、多语种文本检测|[ch_PP-OCRv2_det_cml.yml](../../configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml)| 3M |[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_slim_quant_infer.tar)| |ch_PP-OCRv2_det_slim| slim量化+蒸馏版超轻量模型,支持中英文、多语种文本检测|[ch_PP-OCRv2_det_cml.yml](../../configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml)| 3M |[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_slim_quant_infer.tar)|
|ch_PP-OCRv2_det| 原始超轻量模型,支持中英文、多语种文本检测|[ch_PP-OCRv2_det_cml.yml](../../configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml)|3M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_distill_train.tar)| |ch_PP-OCRv2_det| 原始超轻量模型,支持中英文、多语种文本检测|[ch_PP-OCRv2_det_cml.yml](../../configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml)|3M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_distill_train.tar)|
...@@ -55,7 +55,7 @@ PaddleOCR提供的可下载模型包括`推理模型`、`训练模型`、`预训 ...@@ -55,7 +55,7 @@ PaddleOCR提供的可下载模型包括`推理模型`、`训练模型`、`预训
|模型名称|模型简介|配置文件|推理模型大小|下载地址| |模型名称|模型简介|配置文件|推理模型大小|下载地址|
| --- | --- | --- | --- | --- | | --- | --- | --- | --- | --- |
|en_PP-OCRv3_det_slim |【最新】slim量化版超轻量模型,支持英文、数字检测 | [ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml) | 1.1M |[推理模型(coming soon)](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_slim_infer.tar) / [训练模型(coming soon)](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_slim_distill_train.tar) / [slim模型(coming soon)](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_slim_infer.nb) | |en_PP-OCRv3_det_slim |【最新】slim量化版超轻量模型,支持英文、数字检测 | [ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml) | 1.1M |[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_slim_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_slim_distill_train.tar) / [slim模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_slim_infer.nb) |
|ch_PP-OCRv3_det |【最新】原始超轻量模型,支持英文、数字检测|[ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml)| 3.8M | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_distill_train.tar) | |ch_PP-OCRv3_det |【最新】原始超轻量模型,支持英文、数字检测|[ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml)| 3.8M | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_distill_train.tar) |
* 注:英文检测模型与中文检测模型结构完全相同,只有训练数据不同,在此仅提供相同的配置文件。 * 注:英文检测模型与中文检测模型结构完全相同,只有训练数据不同,在此仅提供相同的配置文件。
...@@ -66,7 +66,7 @@ PaddleOCR提供的可下载模型包括`推理模型`、`训练模型`、`预训 ...@@ -66,7 +66,7 @@ PaddleOCR提供的可下载模型包括`推理模型`、`训练模型`、`预训
|模型名称|模型简介|配置文件|推理模型大小|下载地址| |模型名称|模型简介|配置文件|推理模型大小|下载地址|
| --- | --- | --- | --- | --- | | --- | --- | --- | --- | --- |
| ml_PP-OCRv3_det_slim |【最新】slim量化版超轻量模型,支持多语言检测 | [ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml) | 1.1M |[推理模型(coming soon)](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_slim_infer.tar) / [训练模型(coming soon)](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_slim_distill_train.tar) / [slim模型(coming soon)](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_slim_infer.nb) | | ml_PP-OCRv3_det_slim |【最新】slim量化版超轻量模型,支持多语言检测 | [ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml) | 1.1M |[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_slim_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_slim_distill_train.tar) / [slim模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_slim_infer.nb) |
| ml_PP-OCRv3_det |【最新】原始超轻量模型,支持多语言检测 | [ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml)| 3.8M | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_distill_train.tar) | | ml_PP-OCRv3_det |【最新】原始超轻量模型,支持多语言检测 | [ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml)| 3.8M | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_distill_train.tar) |
* 注:多语言检测模型与中文检测模型结构完全相同,只有训练数据不同,在此仅提供相同的配置文件。 * 注:多语言检测模型与中文检测模型结构完全相同,只有训练数据不同,在此仅提供相同的配置文件。
...@@ -113,11 +113,10 @@ PaddleOCR提供的可下载模型包括`推理模型`、`训练模型`、`预训 ...@@ -113,11 +113,10 @@ PaddleOCR提供的可下载模型包括`推理模型`、`训练模型`、`预训
| te_PP-OCRv3_rec | ppocr/utils/dict/te_dict.txt | 泰卢固文识别|[te_PP-OCRv3_rec.yml](../../configs/rec/PP-OCRv3/multi_language/te_PP-OCRv3_rec.yml)|9.6M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/te_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/te_PP-OCRv3_rec_train.tar) | | te_PP-OCRv3_rec | ppocr/utils/dict/te_dict.txt | 泰卢固文识别|[te_PP-OCRv3_rec.yml](../../configs/rec/PP-OCRv3/multi_language/te_PP-OCRv3_rec.yml)|9.6M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/te_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/te_PP-OCRv3_rec_train.tar) |
| ka_PP-OCRv3_rec | ppocr/utils/dict/ka_dict.txt |卡纳达文识别|[ka_PP-OCRv3_rec.yml](../../configs/rec/PP-OCRv3/multi_language/ka_PP-OCRv3_rec.yml)|9.9M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/ka_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/ka_PP-OCRv3_rec_train.tar) | | ka_PP-OCRv3_rec | ppocr/utils/dict/ka_dict.txt |卡纳达文识别|[ka_PP-OCRv3_rec.yml](../../configs/rec/PP-OCRv3/multi_language/ka_PP-OCRv3_rec.yml)|9.9M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/ka_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/ka_PP-OCRv3_rec_train.tar) |
| ta_PP-OCRv3_rec | ppocr/utils/dict/ta_dict.txt |泰米尔文识别|[ta_PP-OCRv3_rec.yml](../../configs/rec/PP-OCRv3/multi_language/ta_PP-OCRv3_rec.yml)|9.6M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/ta_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/ta_PP-OCRv3_rec_train.tar) | | ta_PP-OCRv3_rec | ppocr/utils/dict/ta_dict.txt |泰米尔文识别|[ta_PP-OCRv3_rec.yml](../../configs/rec/PP-OCRv3/multi_language/ta_PP-OCRv3_rec.yml)|9.6M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/ta_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/ta_PP-OCRv3_rec_train.tar) |
| latin_PP-OCRv3_rec | ppocr/utils/dict/latin_dict.txt | 拉丁文识别 | [latin_PP-OCRv3_rec.yml](../../configs/rec/PP-OCRv3/multi_language/latin_PP-OCRv3_rec.yml) |9.6M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/latin_ppocr_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/latin_ppocr_PP-OCRv3_rec_train.tar) | | latin_PP-OCRv3_rec | ppocr/utils/dict/latin_dict.txt | 拉丁文识别 | [latin_PP-OCRv3_rec.yml](../../configs/rec/PP-OCRv3/multi_language/latin_PP-OCRv3_rec.yml) |9.7M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/latin_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/latin_PP-OCRv3_rec_train.tar) |
| arabic_PP-OCRv3_rec | ppocr/utils/dict/arabic_dict.txt | 阿拉伯字母 | [arabic_PP-OCRv3_rec.yml](../../configs/rec/PP-OCRv3/multi_language/rec_arabic_lite_train.yml) |9.6M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/arabic_ppocr_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/arabic_ppocr_PP-OCRv3_rec_train.tar) | | arabic_PP-OCRv3_rec | ppocr/utils/dict/arabic_dict.txt | 阿拉伯字母 | [arabic_PP-OCRv3_rec.yml](../../configs/rec/PP-OCRv3/multi_language/rec_arabic_lite_train.yml) |9.6M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/arabic_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/arabic_PP-OCRv3_rec_train.tar) |
| cyrillic_PP-OCRv3_rec | ppocr/utils/dict/cyrillic_dict.txt | 斯拉夫字母 | [cyrillic_PP-OCRv3_rec.yml](../../configs/rec/PP-OCRv3/multi_language/cyrillic_PP-OCRv3_rec.yml) |9.6M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/cyrillic_ppocr_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/cyrillic_ppocr_PP-OCRv3_rec_train.tar) | | cyrillic_PP-OCRv3_rec | ppocr/utils/dict/cyrillic_dict.txt | 斯拉夫字母 | [cyrillic_PP-OCRv3_rec.yml](../../configs/rec/PP-OCRv3/multi_language/cyrillic_PP-OCRv3_rec.yml) |9.6M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/cyrillic_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/cyrillic_PP-OCRv3_rec_train.tar) |
| devanagari_PP-OCRv3_rec | ppocr/utils/dict/devanagari_dict.txt |梵文字母 | [devanagari_PP-OCRv3_rec.yml](../../configs/rec/PP-OCRv3/multi_language/devanagari_PP-OCRv3_rec.yml) |9.6M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/devanagari_ppocr_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/devanagari_ppocr_PP-OCRv3_rec_train.tar) | | devanagari_PP-OCRv3_rec | ppocr/utils/dict/devanagari_dict.txt |梵文字母 | [devanagari_PP-OCRv3_rec.yml](../../configs/rec/PP-OCRv3/multi_language/devanagari_PP-OCRv3_rec.yml) |9.9M|[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/devanagari_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/devanagari_PP-OCRv3_rec_train.tar) |
更多支持语种请参考: [多语言模型](./multi_languages.md) 更多支持语种请参考: [多语言模型](./multi_languages.md)
......
...@@ -37,7 +37,7 @@ Relationship of the above models is as follows. ...@@ -37,7 +37,7 @@ Relationship of the above models is as follows.
|model name|description|config|model size|download| |model name|description|config|model size|download|
| --- | --- | --- | --- | --- | | --- | --- | --- | --- | --- |
|ch_PP-OCRv3_det_slim| [New] slim quantization with distillation lightweight model, supporting Chinese, English, multilingual text detection |[ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml)| 1.1M |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_infer.tar) / [trained model (coming soon)](https://paddleocr.bj.bcebos.com/PP-OCRv3/ch/ch_PP-OCRv3_det_slim_distill_train.tar) / [slim model](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_infer.nb)| |ch_PP-OCRv3_det_slim| [New] slim quantization with distillation lightweight model, supporting Chinese, English, multilingual text detection |[ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml)| 1.1M |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv3/ch/ch_PP-OCRv3_det_slim_distill_train.tar) / [slim model](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_infer.nb)|
|ch_PP-OCRv3_det| [New] Original lightweight model, supporting Chinese, English, multilingual text detection |[ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml)| 3.8M |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_distill_train.tar)| |ch_PP-OCRv3_det| [New] Original lightweight model, supporting Chinese, English, multilingual text detection |[ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml)| 3.8M |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_distill_train.tar)|
|ch_PP-OCRv2_det_slim| [New] slim quantization with distillation lightweight model, supporting Chinese, English, multilingual text detection|[ch_PP-OCRv2_det_cml.yml](../../configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml)| 3M |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_slim_quant_infer.tar)| |ch_PP-OCRv2_det_slim| [New] slim quantization with distillation lightweight model, supporting Chinese, English, multilingual text detection|[ch_PP-OCRv2_det_cml.yml](../../configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml)| 3M |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_slim_quant_infer.tar)|
|ch_PP-OCRv2_det| [New] Original lightweight model, supporting Chinese, English, multilingual text detection|[ch_PP-OCRv2_det_cml.yml](../../configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml)|3M|[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_distill_train.tar)| |ch_PP-OCRv2_det| [New] Original lightweight model, supporting Chinese, English, multilingual text detection|[ch_PP-OCRv2_det_cml.yml](../../configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml)|3M|[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_distill_train.tar)|
...@@ -51,7 +51,7 @@ Relationship of the above models is as follows. ...@@ -51,7 +51,7 @@ Relationship of the above models is as follows.
|model name|description|config|model size|download| |model name|description|config|model size|download|
| --- | --- | --- | --- | --- | | --- | --- | --- | --- | --- |
|en_PP-OCRv3_det_slim | [New] Slim qunatization with distillation lightweight detection model, supporting English | [ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml) | 1.1M |[inference model(coming soon)](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_slim_infer.tar) / [trained model (coming soon)](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_slim_distill_train.tar) / [slim model(coming soon)](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_slim_infer.nb) | |en_PP-OCRv3_det_slim | [New] Slim qunatization with distillation lightweight detection model, supporting English | [ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml) | 1.1M |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_slim_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_slim_distill_train.tar) / [slim model](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_slim_infer.nb) |
|ch_PP-OCRv3_det | [New] Original lightweight detection model, supporting English |[ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml)| 3.8M | [inference model](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_distill_train.tar) | |ch_PP-OCRv3_det | [New] Original lightweight detection model, supporting English |[ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml)| 3.8M | [inference model](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_distill_train.tar) |
* Note: English configuration file is same as Chinese except training data, here we only provide one configuration file. * Note: English configuration file is same as Chinese except training data, here we only provide one configuration file.
...@@ -62,7 +62,7 @@ Relationship of the above models is as follows. ...@@ -62,7 +62,7 @@ Relationship of the above models is as follows.
|model name|description|config|model size|download| |model name|description|config|model size|download|
| --- | --- | --- | --- | --- | | --- | --- | --- | --- | --- |
| ml_PP-OCRv3_det_slim | [New] Slim qunatization with distillation lightweight detection model, supporting English | [ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml) | 1.1M | [inference model(coming soon)](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_slim_infer.tar) / [trained model (coming soon)](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_slim_distill_train.tar) / [slim model(coming soon)](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_slim_infer.nb) | | ml_PP-OCRv3_det_slim | [New] Slim qunatization with distillation lightweight detection model, supporting English | [ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml) | 1.1M | [inference model](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_slim_infer.tar) / [trained model ](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_slim_distill_train.tar) / [slim model](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_slim_infer.nb) |
| ml_PP-OCRv3_det |[New] Original lightweight detection model, supporting English | [ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml)| 3.8M | [inference model](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_distill_train.tar) | | ml_PP-OCRv3_det |[New] Original lightweight detection model, supporting English | [ch_PP-OCRv3_det_cml.yml](../../configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml)| 3.8M | [inference model](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv3/multilingual/Multilingual_PP-OCRv3_det_distill_train.tar) |
* Note: English configuration file is same as Chinese except training data, here we only provide one configuration file. * Note: English configuration file is same as Chinese except training data, here we only provide one configuration file.
......
- [PaddleOCR Quick Start](#paddleocr-quick-start) # PaddleOCR Quick Start
- [1. Installation](#1-installation)
**Note:** this tutorial mainly introduces the usage of PP-OCR series models, please refer to [PP-Structure Quick Start](../../ppstructure/docs/quickstart_en.md) for the quick use of document analysis related functions.
- [1. Installation](#1-installation)
- [1.1 Install PaddlePaddle](#11-install-paddlepaddle) - [1.1 Install PaddlePaddle](#11-install-paddlepaddle)
- [1.2 Install PaddleOCR Whl Package](#12-install-paddleocr-whl-package) - [1.2 Install PaddleOCR Whl Package](#12-install-paddleocr-whl-package)
- [2. Easy-to-Use](#2-easy-to-use) - [2. Easy-to-Use](#2-easy-to-use)
- [2.1 Use by Command Line](#21-use-by-command-line) - [2.1 Use by Command Line](#21-use-by-command-line)
- [2.1.1 Chinese and English Model](#211-chinese-and-english-model) - [2.1.1 Chinese and English Model](#211-chinese-and-english-model)
- [2.1.2 Multi-language Model](#212-multi-language-model) - [2.1.2 Multi-language Model](#212-multi-language-model)
...@@ -10,9 +13,8 @@ ...@@ -10,9 +13,8 @@
- [2.2 Use by Code](#22-use-by-code) - [2.2 Use by Code](#22-use-by-code)
- [2.2.1 Chinese & English Model and Multilingual Model](#221-chinese--english-model-and-multilingual-model) - [2.2.1 Chinese & English Model and Multilingual Model](#221-chinese--english-model-and-multilingual-model)
- [2.2.2 Layout Analysis](#222-layout-analysis) - [2.2.2 Layout Analysis](#222-layout-analysis)
- [3. Summary](#3-summary) - [3. Summary](#3-summary)
# PaddleOCR Quick Start
<a name="1nstallation"></a> <a name="1nstallation"></a>
......
doc/ppocr_v3/GTC.png

415.0 KB | W: | H:

doc/ppocr_v3/GTC.png

238.2 KB | W: | H:

doc/ppocr_v3/GTC.png
doc/ppocr_v3/GTC.png
doc/ppocr_v3/GTC.png
doc/ppocr_v3/GTC.png
  • 2-up
  • Swipe
  • Onion skin
doc/ppocr_v3/LKPAN.png

130.2 KB | W: | H:

doc/ppocr_v3/LKPAN.png

92.5 KB | W: | H:

doc/ppocr_v3/LKPAN.png
doc/ppocr_v3/LKPAN.png
doc/ppocr_v3/LKPAN.png
doc/ppocr_v3/LKPAN.png
  • 2-up
  • Swipe
  • Onion skin
doc/ppocr_v3/RSEFPN.png

125.7 KB | W: | H:

doc/ppocr_v3/RSEFPN.png

97.6 KB | W: | H:

doc/ppocr_v3/RSEFPN.png
doc/ppocr_v3/RSEFPN.png
doc/ppocr_v3/RSEFPN.png
doc/ppocr_v3/RSEFPN.png
  • 2-up
  • Swipe
  • Onion skin
doc/ppocr_v3/svtr_g2.png

323.3 KB | W: | H:

doc/ppocr_v3/svtr_g2.png

187.3 KB | W: | H:

doc/ppocr_v3/svtr_g2.png
doc/ppocr_v3/svtr_g2.png
doc/ppocr_v3/svtr_g2.png
doc/ppocr_v3/svtr_g2.png
  • 2-up
  • Swipe
  • Onion skin
doc/ppocr_v3/svtr_g4.png

549.7 KB | W: | H:

doc/ppocr_v3/svtr_g4.png

213.6 KB | W: | H:

doc/ppocr_v3/svtr_g4.png
doc/ppocr_v3/svtr_g4.png
doc/ppocr_v3/svtr_g4.png
doc/ppocr_v3/svtr_g4.png
  • 2-up
  • Swipe
  • Onion skin
doc/ppocr_v3/svtr_tiny.png

585.6 KB | W: | H:

doc/ppocr_v3/svtr_tiny.png

725.7 KB | W: | H:

doc/ppocr_v3/svtr_tiny.png
doc/ppocr_v3/svtr_tiny.png
doc/ppocr_v3/svtr_tiny.png
doc/ppocr_v3/svtr_tiny.png
  • 2-up
  • Swipe
  • Onion skin
doc/ppocrv3_framework.png

957.2 KB | W: | H:

doc/ppocrv3_framework.png

1.1 MB | W: | H:

doc/ppocrv3_framework.png
doc/ppocrv3_framework.png
doc/ppocrv3_framework.png
doc/ppocrv3_framework.png
  • 2-up
  • Swipe
  • Onion skin
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册