Refer to [DTRB](https://arxiv.org/abs/1904.01906), the training and evaluation result of these above text recognition (using MJSynth and SynthText for training, evaluate on IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE) is as follow:
Refer to [DTRB](https://arxiv.org/abs/1904.01906), the training and evaluation result of these above text recognition (using MJSynth and SynthText for training, evaluate on IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE) is as follow:
...
@@ -92,6 +93,7 @@ Refer to [DTRB](https://arxiv.org/abs/1904.01906), the training and evaluation r
...
@@ -92,6 +93,7 @@ Refer to [DTRB](https://arxiv.org/abs/1904.01906), the training and evaluation r
-[3. Model Training / Evaluation / Prediction](#3)
-[3.1 Training](#3-1)
-[3.2 Evaluation](#3-2)
-[3.3 Prediction](#3-3)
-[4. Inference and Deployment](#4)
-[4.1 Python Inference](#4-1)
-[4.2 C++ Inference](#4-2)
-[4.3 Serving](#4-3)
-[4.4 More](#4-4)
-[5. FAQ](#5)
<aname="1"></a>
## 1. Introduction
Paper:
> [RobustScanner: Dynamically Enhancing Positional Clues for Robust Text Recognition](https://arxiv.org/pdf/2007.07542.pdf)
> Xiaoyu Yue, Zhanghui Kuang, Chenhao Lin, Hongbin Sun, Wayne
Zhang
> ECCV, 2020
Using MJSynth and SynthText two text recognition datasets for training, and evaluating on IIIT, SVT, IC13, IC15, SVTP, CUTE datasets, the algorithm reproduction effect is as follows:
Note:In addition to using the two text recognition datasets MJSynth and SynthText, [SynthAdd](https://pan.baidu.com/share/init?surl=uV0LtoNmcxbO-0YA7Ch4dg) data (extraction code: 627x), and some real data are used in training, the specific data details can refer to the paper.
<aname="2"></a>
## 2. Environment
Please refer to ["Environment Preparation"](./environment_en.md) to configure the PaddleOCR environment, and refer to ["Project Clone"](./clone_en.md) to clone the project code.
<aname="3"></a>
## 3. Model Training / Evaluation / Prediction
Please refer to [Text Recognition Tutorial](./recognition_en.md). PaddleOCR modularizes the code, and training different recognition models only requires **changing the configuration file**.
Training:
Specifically, after the data preparation is completed, the training can be started. The training command is as follows:
```
#Single GPU training (long training period, not recommended)
First, the model saved during the RobustScanner text recognition training process is converted into an inference model. you can use the following command to convert:
The following figure shows the result, with different colored detection boxes representing different categories and displaying specific categories in the upper left corner of the box with `show_element_type`
| config_path | model config path | None | Specify config_ path will automatically download the model (only for the first time,the model will exist and will not be downloaded again) |
| label_map | category mapping table | None | Setting config_ path, it can be none, and the label is automatically obtained according to the dataset name_ map, You need to specify it manually when setting model_path |
| enforce_cpu | whether to use CPU | False | False to use GPU, and True to force the use of CPU |
| enforce_mkldnn | whether mkldnn acceleration is enabled in CPU prediction | True | \ |
| thread_num | the number of CPU threads | 10 | \ |
The following model configurations and label maps are currently supported, which you can use by modifying '--config_path' and '--label_map' to detect different types of content:
* TableBank word and TableBank latex are trained on datasets of word documents and latex documents respectively;
**数据分布：**
* Download TableBank dataset contains both word and latex。
## 3. PostProcess
| File or Folder | Description | num |
| :------------- | :------------- | ------- |
| `train/` | 训练集图片 | 335,703 |
| `val/` | 验证集图片 | 11,245 |
| `test/` | 测试集图片 | 11,405 |
| `train.json` | 训练集标注文件 | - |
| `val.json` | 验证集标注文件 | - |
Layout parser contains multiple categories, if you only want to get the detection box for a specific category (such as the "Text" category), you can use the following code:
PaddleDetection支持了基于FGD([Focal and Global Knowledge Distillation for Detectors](https://arxiv.org/abs/2111.11837v1))蒸馏的目标检测模型训练过程，FGD蒸馏分为两个部分`Focal`和`Global`。`Focal`蒸馏分离图像的前景和背景，让学生模型分别关注教师模型的前景和背景部分特征的关键像素；`Global`蒸馏部分重建不同像素之间的关系并将其从教师转移到学生，以补偿`Focal`蒸馏中丢失的全局信息。
title={PubLayNet: largest dataset ever for document layout analysis},
author={Zhong, Xu and Tang, Jianbin and Yepes, Antonio Jimeno},
booktitle={2019 International Conference on Document Analysis and Recognition (ICDAR)},
year={2019},
volume={},
number={},
pages={1015-1022},
doi={10.1109/ICDAR.2019.00166},
ISSN={1520-5363},
month={Sep.},
organization={IEEE}
}
@inproceedings{yang2022focal,
title={Focal and global knowledge distillation for detectors},
author={Yang, Zhendong and Li, Zhe and Jiang, Xiaohu and Gong, Yuan and Yuan, Zehuan and Zhao, Danpei and Yuan, Chun},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={4643--4652},
year={2022}
}
```
The above model is based on [PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection). If you want to train your own layout parser model，please refer to：[train_layoutparser_model](train_layoutparser_model.md)
For more installation tutorials, please refer to： [Install doc](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.1/docs/tutorials/INSTALL_cn.md)
## 2. Data preparation
Download the [PubLayNet](https://github.com/ibm-aur-nlp/PubLayNet) dataset
| `train/` | Images in the training subset | 335,703 |
| `val/` | Images in the validation subset | 11,245 |
| `test/` | Images in the testing subset | 11,405 |
| `train.json` | Annotations for training images | 1 |
| `val.json` | Annotations for validation images | 1 |
| `LICENSE.txt` | Plaintext version of the CDLA-Permissive license | 1 |
| `README.txt` | Text file with the file names and description | 1 |
For other datasets，please refer to [the PrepareDataSet]((https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.1/docs/tutorials/PrepareDataSet.md) )
## 3. Configuration
We use the `configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml` configuration for training，the configuration file is as follows
The `ppyolov2_r50vd_dcn_365e_coco.yml` configuration depends on other configuration files, in this case:
- coco_detection.yml：mainly explains the path of training data and verification data
- runtime.yml：mainly describes the common parameters, such as whether to use the GPU and how many epoch to save model etc.
- optimizer_365e.yml：mainly explains the learning rate and optimizer configuration
- ppyolov2_r50vd_dcn.yml：mainly describes the model and the network
- ppyolov2_reader.yml：mainly describes the configuration of data readers, such as batch size and number of concurrent loading child processes, and also includes post preprocessing, such as resize and data augmention etc.
Modify the preceding files, such as the dataset path and batch size etc.
## 4. Training
PaddleDetection provides single-card/multi-card training mode to meet various training needs of users:
* GPU single card training
```bash
export CUDA_VISIBLE_DEVICES=0 #Don't need to run this command on Windows and Mac
`--draw_threshold` is an optional parameter. According to the calculation of [NMS](https://ieeexplore.ieee.org/document/1699659), different threshold will produce different results, ` keep_top_k ` represent the maximum amount of output target, the default value is 10. You can set different value according to your own actual situation。
## 6. Deployment
Use your trained model in Layout Parser
### 6.1 Export model
n the process of model training, the model file saved contains the process of forward prediction and back propagation. In the actual industrial deployment, there is no need for back propagation. Therefore, the model should be translated into the model format required by the deployment. The `tools/export_model.py` script is provided in PaddleDetection to export the model.
The exported model name defaults to `model.*`, Layout Parser's code model is `inference.*`, So change [PaddleDetection/ppdet/engine/trainer. Py ](https://github.com/PaddlePaddle/PaddleDetection/blob/b87a1ea86fa18ce69e44a17ad1b49c1326f19ff9/ppdet/engine/trainer.py# L512) (click on the link to see the detailed line of code), change 'model' to 'inference'.
The prediction model is exported to `inference/ppyolov2_r50vd_dcn_365e_coco` ,including:`infer_cfg.yml`(prediction not required), `inference.pdiparams`, `inference.pdiparams.info`,`inference.pdmodel`
More model export tutorials, please refer to：[EXPORT_MODEL](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.1/deploy/EXPORT_MODEL.md)
### 6.2 Inference
`model_path` represent the trained model path, and layoutparser is used to predict:
More PaddleDetection training tutorials，please reference：[PaddleDetection Training](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.1/docs/tutorials/GETTING_STARTED_cn.md)
