recognition_en.md 16.7 KB
Newer Older
T
tink2123 已提交
1
# TEXT RECOGNITION
K
Khanh Tran 已提交
2

W
WenmuZhou 已提交
3 4 5 6 7
- [1 DATA PREPARATION](#DATA_PREPARATION)
    - [1.1 Costom Dataset](#Costom_Dataset)
    - [1.2 Dataset Download](#Dataset_download)
    - [1.3 Dictionary](#Dictionary)  
    - [1.4 Add Space Category](#Add_space_category)
W
WenmuZhou 已提交
8

W
WenmuZhou 已提交
9 10
- [2 TRAINING](#TRAINING)
    - [2.1 Data Augmentation](#Data_Augmentation)
T
tink2123 已提交
11 12
    - [2.2 General Training](#Training)
    - [2.3 Multi-language Training](#Multi_language)
W
WenmuZhou 已提交
13

W
WenmuZhou 已提交
14
- [3 EVALUATION](#EVALUATION)
W
WenmuZhou 已提交
15

W
WenmuZhou 已提交
16 17
- [4 PREDICTION](#PREDICTION)
    - [4.1 Training engine prediction](#Training_engine_prediction)
W
WenmuZhou 已提交
18 19

<a name="DATA_PREPARATION"></a>
T
tink2123 已提交
20
## 1 DATA PREPARATION
K
Khanh Tran 已提交
21 22


W
WenmuZhou 已提交
23
PaddleOCR supports two data formats:
T
tink2123 已提交
24 25
- `LMDB` is used to train data sets stored in lmdb format(LMDBDataSet);
- `general data` is used to train data sets stored in text files(SimpleDataSet):
K
Khanh Tran 已提交
26 27 28 29 30 31

Please organize the dataset as follows:

The default storage path for training data is `PaddleOCR/train_data`, if you already have a dataset on your disk, just create a soft link to the dataset directory:

```
W
WenmuZhou 已提交
32
# linux and mac os
33
ln -sf <path/to/dataset> <path/to/paddle_ocr>/train_data/dataset
W
WenmuZhou 已提交
34 35
# windows
mklink /d <path/to/paddle_ocr>/train_data/dataset <path/to/dataset>
K
Khanh Tran 已提交
36 37
```

W
WenmuZhou 已提交
38
<a name="Costom_Dataset"></a>
T
tink2123 已提交
39
### 1.1 Costom dataset
K
Khanh Tran 已提交
40 41 42 43 44

If you want to use your own data for training, please refer to the following to organize your data.

- Training set

W
WenmuZhou 已提交
45
It is recommended to put the training images in the same folder, and use a txt file (rec_gt_train.txt) to store the image path and label. The contents of the txt file are as follows:
K
Khanh Tran 已提交
46 47 48 49 50 51

* Note: by default, the image path and image label are split with \t, if you use other methods to split, it will cause training error

```
" Image file name           Image annotation "

W
WenmuZhou 已提交
52 53
train_data/rec/train/word_001.jpg   简单可依赖
train_data/rec/train/word_002.jpg   用科技让复杂的世界更简单
W
WenmuZhou 已提交
54
...
K
Khanh Tran 已提交
55 56 57 58 59 60
```

The final training set should have the following file structure:

```
|-train_data
W
WenmuZhou 已提交
61
  |-rec
W
WenmuZhou 已提交
62 63 64 65 66 67
    |- rec_gt_train.txt
    |- train
        |- word_001.png
        |- word_002.jpg
        |- word_003.jpg
        | ...
K
Khanh Tran 已提交
68 69 70 71 72 73 74 75
```

- Test set

Similar to the training set, the test set also needs to be provided a folder containing all images (test) and a rec_gt_test.txt. The structure of the test set is as follows:

```
|-train_data
W
WenmuZhou 已提交
76
  |-rec
K
Khanh Tran 已提交
77 78 79 80 81 82 83 84
    |-ic15_data
        |- rec_gt_test.txt
        |- test
            |- word_001.jpg
            |- word_002.jpg
            |- word_003.jpg
            | ...
```
W
WenmuZhou 已提交
85 86

<a name="Dataset_download"></a>
T
tink2123 已提交
87
### 1.2 Dataset download
W
WenmuZhou 已提交
88

T
tink2123 已提交
89
- ICDAR2015
W
WenmuZhou 已提交
90

T
tink2123 已提交
91 92
If you do not have a dataset locally, you can download it on the official website [icdar2015](http://rrc.cvc.uab.es/?ch=4&com=downloads).
Also refer to [DTRB](https://github.com/clovaai/deep-text-recognition-benchmark#download-lmdb-dataset-for-traininig-and-evaluation-from-here) ,download the lmdb format dataset required for benchmark
W
WenmuZhou 已提交
93 94 95 96 97 98 99 100 101 102

PaddleOCR provides label files for training the icdar2015 dataset, which can be downloaded in the following ways:

```
# Training set label
wget -P ./train_data/ic15_data  https://paddleocr.bj.bcebos.com/dataset/rec_gt_train.txt
# Test Set Label
wget -P ./train_data/ic15_data  https://paddleocr.bj.bcebos.com/dataset/rec_gt_test.txt
```

T
tink2123 已提交
103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122
PaddleOCR also provides a data format conversion script, which can convert ICDAR official website label to a data format
supported by PaddleOCR. The data conversion tool is in `ppocr/utils/gen_label.py`, here is the training set as an example:

```
# convert the official gt to rec_gt_label.txt
python gen_label.py --mode="rec" --input_path="{path/of/origin/label}" --output_label="rec_gt_label.txt"
```

The data format is as follows, (a) is the original picture, (b) is the Ground Truth text file corresponding to each picture:

![](../datasets/icdar_rec.png)


- Multilingual dataset

The multi-language model training method is the same as the Chinese model. The training data set is 100w synthetic data. A small amount of fonts and test data can be downloaded using the following two methods.
* [Baidu Netdisk](https://pan.baidu.com/s/1bS_u207Rm7YbY33wOECKDA) ,Extraction code:frgi.
* [Google drive](https://drive.google.com/file/d/18cSWX7wXSy4G0tbKJ0d9PuIaiwRLHpjA/view)


W
WenmuZhou 已提交
123
<a name="Dictionary"></a>
T
tink2123 已提交
124
### 1.3 Dictionary
K
Khanh Tran 已提交
125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140

Finally, a dictionary ({word_dict_name}.txt) needs to be provided so that when the model is trained, all the characters that appear can be mapped to the dictionary index.

Therefore, the dictionary needs to contain all the characters that you want to be recognized correctly. {word_dict_name}.txt needs to be written in the following format and saved in the `utf-8` encoding format:

```
l
d
a
d
r
n
```

In `word_dict.txt`, there is a single word in each line, which maps characters and numeric indexes together, e.g "and" will be mapped to [2 5 1]

W
WenmuZhou 已提交
141 142
PaddleOCR has built-in dictionaries, which can be used on demand.

K
Khanh Tran 已提交
143 144
`ppocr/utils/ppocr_keys_v1.txt` is a Chinese dictionary with 6623 characters.

W
WenmuZhou 已提交
145 146 147 148
`ppocr/utils/ic15_dict.txt` is an English dictionary with 63 characters

`ppocr/utils/dict/french_dict.txt` is a French dictionary with 118 characters

149
`ppocr/utils/dict/japan_dict.txt` is a Japanese dictionary with 4399 characters
W
WenmuZhou 已提交
150

T
tink2123 已提交
151
`ppocr/utils/dict/korean_dict.txt` is a Korean dictionary with 3636 characters
W
WenmuZhou 已提交
152

T
tink2123 已提交
153 154
`ppocr/utils/dict/german_dict.txt` is a German dictionary with 131 characters

T
tink2123 已提交
155
`ppocr/utils/en_dict.txt` is a English dictionary with 96 characters
W
WenmuZhou 已提交
156

X
xiaoting 已提交
157

W
WenmuZhou 已提交
158
The current multi-language model is still in the demo stage and will continue to optimize the model and add languages. **You are very welcome to provide us with dictionaries and fonts in other languages**,
littletomatodonkey's avatar
fix doc  
littletomatodonkey 已提交
159
If you like, you can submit the dictionary file to [dict](../../ppocr/utils/dict) and we will thank you in the Repo.
K
Khanh Tran 已提交
160 161 162 163


To customize the dict file, please modify the `character_dict_path` field in `configs/rec/rec_icdar15_train.yml` and set `character_type` to `ch`.

T
tink2123 已提交
164 165 166 167
- Custom dictionary

If you need to customize dic file, please add character_dict_path field in configs/rec/rec_icdar15_train.yml to point to your dictionary path. And set character_type to ch.

W
WenmuZhou 已提交
168
<a name="Add_space_category"></a>
T
tink2123 已提交
169
### 1.4 Add space category
T
tink2123 已提交
170

171
If you want to support the recognition of the `space` category, please set the `use_space_char` field in the yml file to `True`.
T
tink2123 已提交
172 173 174

**Note: use_space_char only takes effect when character_type=ch**

W
WenmuZhou 已提交
175
<a name="TRAINING"></a>
T
tink2123 已提交
176
## 2 TRAINING
K
Khanh Tran 已提交
177

T
tink2123 已提交
178
<a name="Data_Augmentation"></a>
T
tink2123 已提交
179
### 2.1 Data Augmentation
T
tink2123 已提交
180 181 182 183 184 185 186 187

PaddleOCR provides a variety of data augmentation methods. All the augmentation methods are enabled by default.

The default perturbation methods are: cvtColor, blur, jitter, Gasuss noise, random crop, perspective, color reverse, TIA augmentation.

Each disturbance method is selected with a 40% probability during the training process. For specific code implementation, please refer to: [rec_img_aug.py](../../ppocr/data/imaug/rec_img_aug.py)

<a name="Training"></a>
T
tink2123 已提交
188
### 2.2 General Training
T
tink2123 已提交
189

K
Khanh Tran 已提交
190 191 192 193 194 195 196
PaddleOCR provides training scripts, evaluation scripts, and prediction scripts. In this section, the CRNN recognition model will be used as an example:

First download the pretrain model, you can download the trained model to finetune on the icdar2015 data:

```
cd PaddleOCR/
# Download the pre-trained model of MobileNetV3
T
tink2123 已提交
197
wget -P ./pretrain_models/ https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mv3_none_bilstm_ctc_v2.0_train.tar
K
Khanh Tran 已提交
198 199
# Decompress model parameters
cd pretrain_models
T
tink2123 已提交
200
tar -xf rec_mv3_none_bilstm_ctc_v2.0_train.tar && rm -rf rec_mv3_none_bilstm_ctc_v2.0_train.tar
K
Khanh Tran 已提交
201 202 203 204 205
```

Start training:

```
T
tink2123 已提交
206
# GPU training Support single card and multi-card training
T
tink2123 已提交
207
# Training icdar15 English data and The training log will be automatically saved as train.log under "{save_model_dir}"
T
tink2123 已提交
208 209 210 211

#specify the single card training(Long training time, not recommended)
python3 tools/train.py -c configs/rec/rec_icdar15_train.yml
#specify the card number through --gpus
212
python3 -m paddle.distributed.launch --gpus '0,1,2,3'  tools/train.py -c configs/rec/rec_icdar15_train.yml
K
Khanh Tran 已提交
213
```
T
tink2123 已提交
214 215


K
Khanh Tran 已提交
216 217 218 219 220 221 222 223 224
PaddleOCR supports alternating training and evaluation. You can modify `eval_batch_step` in `configs/rec/rec_icdar15_train.yml` to set the evaluation frequency. By default, it is evaluated every 500 iter and the best acc model is saved under `output/rec_CRNN/best_accuracy` during the evaluation process.

If the evaluation set is large, the test will be time-consuming. It is recommended to reduce the number of evaluations, or evaluate after training.

* Tip: You can use the `-c` parameter to select multiple model configurations under the `configs/rec/` path for training. The recognition algorithms supported by PaddleOCR are:


| Configuration file |  Algorithm |   backbone |   trans   |   seq      |     pred     |
| :--------: |  :-------:   | :-------:  |   :-------:   |   :-----:   |  :-----:   |
225 226
| [rec_chinese_lite_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml) |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  |
| [rec_chinese_common_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_common_train_v2.0.yml) |  CRNN | ResNet34_vd |  None   |  BiLSTM |  ctc  |
K
Khanh Tran 已提交
227
| rec_chinese_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  |
W
WenmuZhou 已提交
228
| rec_chinese_common_train.yml |  CRNN |   ResNet34_vd |  None   |  BiLSTM |  ctc  |
K
Khanh Tran 已提交
229 230 231 232 233
| rec_icdar15_train.yml |  CRNN |   Mobilenet_v3 large 0.5 |  None   |  BiLSTM |  ctc  |
| rec_mv3_none_bilstm_ctc.yml |  CRNN |   Mobilenet_v3 large 0.5 |  None   |  BiLSTM |  ctc  |
| rec_mv3_none_none_ctc.yml |  Rosetta |   Mobilenet_v3 large 0.5 |  None   |  None |  ctc  |
| rec_r34_vd_none_bilstm_ctc.yml |  CRNN |   Resnet34_vd |  None   |  BiLSTM |  ctc  |
| rec_r34_vd_none_none_ctc.yml |  Rosetta |   Resnet34_vd |  None   |  None |  ctc  |
L
LDOUBLEV 已提交
234 235
| rec_mv3_tps_bilstm_att.yml |  CRNN |   Mobilenet_v3 |  TPS   |  BiLSTM |  att  |
| rec_r34_vd_tps_bilstm_att.yml |  CRNN |   Resnet34_vd |  TPS   |  BiLSTM |  att  |
T
tink2123 已提交
236
| rec_r50fpn_vd_none_srn.yml    | SRN | Resnet50_fpn_vd    | None    | rnn | srn |
T
Topdu 已提交
237
| rec_mtb_nrtr.yml    | NRTR | nrtr_mtb    | None    | transformer encoder | transformer decoder |
K
Khanh Tran 已提交
238

W
WenmuZhou 已提交
239
For training Chinese data, it is recommended to use
240
[rec_chinese_lite_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml). If you want to try the result of other algorithms on the Chinese data set, please refer to the following instructions to modify the configuration file:
K
Khanh Tran 已提交
241
co
242
Take `rec_chinese_lite_train_v2.0.yml` as an example:
K
Khanh Tran 已提交
243 244 245
```
Global:
  ...
246 247
  # Add a custom dictionary, such as modify the dictionary, please point the path to the new dictionary
  character_dict_path: ppocr/utils/ppocr_keys_v1.txt
K
Khanh Tran 已提交
248 249 250
  # Modify character type
  character_type: ch
  ...
251
  # Whether to recognize spaces
252
  use_space_char: True
K
Khanh Tran 已提交
253

254 255 256 257

Optimizer:
  ...
  # Add learning rate decay strategy
258 259 260 261 262 263 264 265 266
  lr:
    name: Cosine
    learning_rate: 0.001
  ...

...

Train:
  dataset:
M
MissPenguin 已提交
267
    # Type of dataset,we support LMDBDataSet and SimpleDataSet
268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286
    name: SimpleDataSet
    # Path of dataset
    data_dir: ./train_data/
    # Path of train list
    label_file_list: ["./train_data/train_list.txt"]
    transforms:
      ...
      - RecResizeImg:
          # Modify image_shape to fit long text
          image_shape: [3, 32, 320]
      ...
  loader:
    ...
    # Train batch_size for Single card
    batch_size_per_card: 256
    ...

Eval:
  dataset:
M
MissPenguin 已提交
287
    # Type of dataset,we support LMDBDataSet and SimpleDataSet
288 289 290 291 292 293 294 295 296 297 298 299 300 301 302
    name: SimpleDataSet
    # Path of dataset
    data_dir: ./train_data
    # Path of eval list
    label_file_list: ["./train_data/val_list.txt"]
    transforms:
      ...
      - RecResizeImg:
          # Modify image_shape to fit long text
          image_shape: [3, 32, 320]
      ...
  loader:
    # Eval batch_size for Single card
    batch_size_per_card: 256
    ...
K
Khanh Tran 已提交
303 304 305
```
**Note that the configuration file for prediction/evaluation must be consistent with the training.**

W
WenmuZhou 已提交
306
<a name="Multi_language"></a>
T
tink2123 已提交
307
### 2.3 Multi-language Training
T
tink2123 已提交
308 309 310

Currently, the multi-language algorithms supported by PaddleOCR are:

T
tink2123 已提交
311
| Configuration file |  Algorithm name |   backbone |   trans   |   seq      |     pred     |  language | character_type |
T
tink2123 已提交
312 313
| :--------: |  :-------:   | :-------:  |   :-------:   |   :-----:   |  :-----:   | :-----:  | :-----:  |
| rec_chinese_cht_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | chinese traditional  | chinese_cht|
T
tink2123 已提交
314
| rec_en_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | English(Case sensitive)   | EN |
T
tink2123 已提交
315 316 317 318
| rec_french_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | French |  french |
| rec_ger_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | German   | german |
| rec_japan_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | Japanese | japan |
| rec_korean_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | Korean  | korean |
T
tink2123 已提交
319 320 321 322
| rec_latin_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | Latin  | latin |
| rec_arabic_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | arabic |  ar |
| rec_cyrillic_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | cyrillic   | cyrillic |
| rec_devanagari_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | devanagari  | devanagari |
T
tink2123 已提交
323

T
tink2123 已提交
324
For more supported languages, please refer to : [Multi-language model](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.1/doc/doc_en/multi_languages_en.md#4-support-languages-and-abbreviations)
W
WenmuZhou 已提交
325 326 327 328 329 330 331 332 333


If you want to finetune on the basis of the existing model effect, please refer to the following instructions to modify the configuration file:

Take `rec_french_lite_train` as an example:

```
Global:
  ...
334
  # Add a custom dictionary, such as modify the dictionary, please point the path to the new dictionary
W
WenmuZhou 已提交
335 336
  character_dict_path: ./ppocr/utils/dict/french_dict.txt
  ...
337
  # Whether to recognize spaces
338
  use_space_char: True
339

W
WenmuZhou 已提交
340
...
341 342 343

Train:
  dataset:
M
MissPenguin 已提交
344
    # Type of dataset,we support LMDBDataSet and SimpleDataSet
345 346 347 348 349 350 351 352 353
    name: SimpleDataSet
    # Path of dataset
    data_dir: ./train_data/
    # Path of train list
    label_file_list: ["./train_data/french_train.txt"]
    ...

Eval:
  dataset:
M
MissPenguin 已提交
354
    # Type of dataset,we support LMDBDataSet and SimpleDataSet
355 356 357 358 359 360
    name: SimpleDataSet
    # Path of dataset
    data_dir: ./train_data
    # Path of eval list
    label_file_list: ["./train_data/french_val.txt"]
    ...
W
WenmuZhou 已提交
361
```
K
Khanh Tran 已提交
362

W
WenmuZhou 已提交
363
<a name="EVALUATION"></a>
T
tink2123 已提交
364
## 3 EVALUATION
K
Khanh Tran 已提交
365

W
WenmuZhou 已提交
366
The evaluation dataset can be set by modifying the `Eval.dataset.label_file_list` field in the `configs/rec/rec_icdar15_train.yml` file.
K
Khanh Tran 已提交
367 368 369

```
# GPU evaluation, Global.checkpoints is the weight to be tested
W
WenmuZhou 已提交
370
python3 -m paddle.distributed.launch --gpus '0' tools/eval.py -c configs/rec/rec_icdar15_train.yml -o Global.checkpoints={path/to/weights}/best_accuracy
K
Khanh Tran 已提交
371 372
```

W
WenmuZhou 已提交
373
<a name="PREDICTION"></a>
T
tink2123 已提交
374
## 4 PREDICTION
K
Khanh Tran 已提交
375 376 377 378


Using the model trained by paddleocr, you can quickly get prediction through the following script.

T
tink2123 已提交
379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399
The default prediction picture is stored in `infer_img`, and the trained weight is specified via `-o Global.checkpoints`:


According to the `save_model_dir` and `save_epoch_step` fields set in the configuration file, the following parameters will be saved:

```
output/rec/
├── best_accuracy.pdopt  
├── best_accuracy.pdparams  
├── best_accuracy.states  
├── config.yml  
├── iter_epoch_3.pdopt  
├── iter_epoch_3.pdparams  
├── iter_epoch_3.states  
├── latest.pdopt  
├── latest.pdparams  
├── latest.states  
└── train.log
```

Among them, best_accuracy.* is the best model on the evaluation set; iter_epoch_x.* is the model saved at intervals of `save_epoch_step`; latest.* is the model of the last epoch.
K
Khanh Tran 已提交
400 401 402

```
# Predict English results
W
WenmuZhou 已提交
403
python3 tools/infer_rec.py -c configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml -o Global.pretrained_model={path/to/weights}/best_accuracy Global.load_static_weights=false Global.infer_img=doc/imgs_words/en/word_1.jpg
K
Khanh Tran 已提交
404 405
```

T
tink2123 已提交
406

K
Khanh Tran 已提交
407 408
Input image:

409
![](../imgs_words/en/word_1.png)
K
Khanh Tran 已提交
410 411 412 413 414

Get the prediction result of the input image:

```
infer_img: doc/imgs_words/en/word_1.png
T
tink2123 已提交
415
        result: ('joint', 0.9998967)
K
Khanh Tran 已提交
416 417
```

418
The configuration file used for prediction must be consistent with the training. For example, you completed the training of the Chinese model with `python3 tools/train.py -c configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml`, you can use the following command to predict the Chinese model:
K
Khanh Tran 已提交
419 420 421

```
# Predict Chinese results
W
WenmuZhou 已提交
422
python3 tools/infer_rec.py -c configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml -o Global.pretrained_model={path/to/weights}/best_accuracy Global.load_static_weights=false Global.infer_img=doc/imgs_words/ch/word_1.jpg
K
Khanh Tran 已提交
423 424 425 426
```

Input image:

427
![](../imgs_words/ch/word_1.jpg)
K
Khanh Tran 已提交
428 429 430 431 432

Get the prediction result of the input image:

```
infer_img: doc/imgs_words/ch/word_1.jpg
T
tink2123 已提交
433
        result: ('韩国小馆', 0.997218)
K
Khanh Tran 已提交
434
```