README.md 4.55 KB
Newer Older
sugon_cxj's avatar
sugon_cxj committed
1
# PaddleOCR
chenxj's avatar
chenxj committed
2
## 论文
chenxj's avatar
chenxj committed
3
PaddleOCR通过det、rec、cls三个模型分别实现字符检测、字符识别和字符方向分类的应用
chenxj's avatar
chenxj committed
4

chenxj's avatar
chenxj committed
5
6
7
8
9
10
11
12
13
det模型主要用DB算法,参考论文如下:

https://arxiv.org/pdf/1911.08947.pdf

rec模型主要用SVTR算法,参考论文如下:

https://arxiv.org/pdf/2205.00159.pdf

cls模型用mobilenetv3实现通用分类,参考论文如下:
chenxj's avatar
chenxj committed
14

chenxj's avatar
chenxj committed
15
https://arxiv.org/pdf/1905.02244.pdf
sugon_cxj's avatar
sugon_cxj committed
16
## 模型结构
chenxj's avatar
chenxj committed
17
det:
chenxj's avatar
chenxj committed
18
![image](https://developer.hpccube.com/codes/modelzoo/paddleocr/-/raw/main/configs/det/dbnet-arc.png)
chenxj's avatar
chenxj committed
19

chenxj's avatar
chenxj committed
20
rec:
chenxj's avatar
chenxj committed
21
![image](https://developer.hpccube.com/codes/modelzoo/paddleocr/-/raw/main/configs/rec/SVTR-arc.png)
chenxj's avatar
chenxj committed
22

chenxj's avatar
chenxj committed
23
cls:
chenxj's avatar
chenxj committed
24
![image](https://developer.hpccube.com/codes/modelzoo/paddleocr/-/raw/main/configs/cls/mobilenetv3-arc.png)
chenxj's avatar
chenxj committed
25
26
## 算法原理
det->cls->rec->text
sugon_cxj's avatar
sugon_cxj committed
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
## 数据集
推荐使用icdar2015数据集[icdar2015](https://rrc.cvc.uab.es/?ch=4&com=downloads)

检测模型训练集文件结构
```
/PaddleOCR/train_data/icdar2015/text_localization/
  └─ icdar_c4_train_imgs/         Training data of icdar dataset
  └─ ch4_test_images/             Testing data of icdar dataset
  └─ train_icdar2015_label.txt    Training annotation of icdar dataset
  └─ test_icdar2015_label.txt     Test annotation of icdar dataset
```
识别模型训练集文件结构
```
|-train_data
  |-rec
    |- rec_gt_train.txt
    |- train
        |- word_001.png
        |- word_002.jpg
        |- word_003.jpg
        | ...
    |-ic15_data
        |- rec_gt_test.txt
        |- test
            |- word_001.jpg
            |- word_002.jpg
            |- word_003.jpg
            | ...
```
chenxj's avatar
chenxj committed
56
## 环境配置
sugon_cxj's avatar
sugon_cxj committed
57
[光源](https://www.sourcefind.cn/#/service-details)可拉取训练以及推理的docker镜像,在[光合开发者社区](https://cancon.hpccube.com:65024/4/main/)可下载paddle安装包。PaddleOCR推荐的镜像如下:
chenxj's avatar
chenxj committed
58
59
60
```
docker pull image.sourcefind.cn:5000/dcu/admin/base/paddlepaddle:2.3.2-centos7.6-dtk-22.10.1-py37-latest
```
chenxj's avatar
chenxj committed
61
## 训练
sugon_cxj's avatar
sugon_cxj committed
62
63

检测模型
chenxj's avatar
chenxj committed
64
```
sugon_cxj's avatar
sugon_cxj committed
65
python3 -m paddle.distributed.launch --gpus '0,1,2,3'  tools/train.py -c configs/det/det_mv3_db.yml -o Global.pretrained_model=./pretrain_models/MobileNetV3_large_x0_5_pretrained
chenxj's avatar
chenxj committed
66
```
sugon_cxj's avatar
sugon_cxj committed
67
识别模型
chenxj's avatar
chenxj committed
68
```
sugon_cxj's avatar
sugon_cxj committed
69
python3 -m paddle.distributed.launch --gpus '0,1,2,3'  tools/train.py -c configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml -o Global.pretrained_model=./pretrain_models/en_PP-OCRv3_rec_train/best_accuracy
chenxj's avatar
chenxj committed
70
```
chenxj's avatar
chenxj committed
71
## 测试
sugon_cxj's avatar
sugon_cxj committed
72
检测模型
chenxj's avatar
chenxj committed
73
```
sugon_cxj's avatar
sugon_cxj committed
74
python3 -m paddle.distributed.launch --gpus '0,1,2,3'  tools/eval.py -c configs/det/det_mv3_db.yml -o Global.pretrained_model=./output/db_mv3/best_accuracy.pdparams
chenxj's avatar
chenxj committed
75
```
sugon_cxj's avatar
sugon_cxj committed
76
识别模型
chenxj's avatar
chenxj committed
77
```
sugon_cxj's avatar
sugon_cxj committed
78
python3 -m paddle.distributed.launch --gpus '0,1,2,3'  tools/eval.py -c configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml -o Global.pretrained_model=./output/v3_en_mobile/best_accuracy.pdparams
chenxj's avatar
chenxj committed
79
```
chenxj's avatar
chenxj committed
80
## 测试(ort)
chenxj's avatar
chenxj committed
81
82
83
84
85
86
87
88
检测模型
```
python3 -m paddle.distributed.launch --gpus '0,1,2,3'  tools/eval.py -c configs/det/det_mv3_db.yml -o Global.pretrained_model=./ch_PP-OCRv3_det_infer/ch_PP-OCRv3_det.onnx --use_onnx=true
```
识别模型
```
python3 -m paddle.distributed.launch --gpus '0,1,2,3'  tools/eval.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec.yml -o Global.pretrained_model=./ch_PP-OCRv3_rec_infer/ch_PP-OCRv3_rec.onnx --use_onnx=true
```
chenxj's avatar
chenxj committed
89
## 推理
chenxj's avatar
chenxj committed
90
```
sugon_cxj's avatar
sugon_cxj committed
91
python3 tools/infer/predict_system.py --image_dir="./doc/imgs/" --det_model_dir="./ch_PP-OCRv3_det_infer/" --rec_model_dir="./ch_PP-OCRv3_rec_infer/" --use_angle_cls=false --rec_image_shape=3,48,320 --warmup=1
chenxj's avatar
chenxj committed
92
```
chenxj's avatar
chenxj committed
93
## 推理(ort)
94
95
96
```
python3 tools/infer/predict_system.py --image_dir="./doc/imgs/" --det_model_dir="./ch_PP-OCRv3_det_infer/ch_PP-OCRv3_det.onnx" --cls_model_dir="./ch_ppocr_mobile_v2.0_cls_infer/ch_ppocr_mobile_v2.0_cls_infer.onnx" --rec_model_dir="./ch_PP-OCRv3_rec_infer/ch_PP-OCRv3_rec.onnx" --use_onnx=true --use_angle_cls=true --rec_image_shape=3,48,320 --warmup=1
```
chenxj's avatar
chenxj committed
97
98
99
## result
![image](https://developer.hpccube.com/codes/modelzoo/paddleocr/-/raw/main/inference_results/08.jpg)
### 性能和准确率数据
sugon_cxj's avatar
sugon_cxj committed
100
101
102
103
104
105
106
107
108
109

检测模型测试
| Model | Precision | Recall |
| :------: | :------: |:------: |
| det | 0.7054 | 0.7193  |

识别模型测试
| Model | Acc | 
| :------: | :------: |
| rec | 0.6490 | 
chenxj's avatar
chenxj committed
110
111
112
113
114
115
116
117
118
119

检测模型测试(ort)
| Model | Precision | Recall |
| :------: | :------: |:------: |
| det | 0.5097 | 0.4068  |

识别模型测试(ort)
| Model | Acc | 
| :------: | :------: |
| rec | 0.6076 | 
chenxj's avatar
chenxj committed
120
121
122
123
124
## 应用场景
### 算法类别
ocr
### 热点应用行业
工业制造、金融、交通、教育、医疗
chenxj's avatar
chenxj committed
125
## 源码仓库及问题反馈
chenxj's avatar
chenxj committed
126
https://developer.hpccube.com/codes/modelzoo/paddleocr
sugon_cxj's avatar
sugon_cxj committed
127
128
129
## 参考
* [PaddleOCR](https://github.com/PaddlePaddle/PaddleOCR)