-`trainValTestRatio` is the division ratio of the number of images in the training set, validation set, and test set, set according to your actual situation, the default is `6:2:2`
-`trainValTestRatio` is the division ratio of the number of images in the training set, validation set, and test set, set according to your actual situation, the default is `6:2:2`
-`labelRootPath` is the storage path of the dataset labeled by PPOCRLabel, the default is `../train_data/label`
-`datasetRootPath` is the storage path of the complete dataset labeled by PPOCRLabel. The default path is `PaddleOCR/train_data` .
```
-`detRootPath` is the path where the text detection dataset is divided according to the dataset marked by PPOCRLabel. The default is `../train_data/det`
|-train_data
|-crop_img
-`recRootPath` is the path where the character recognition dataset is divided according to the dataset marked by PPOCRLabel. The default is `../train_data/rec`
The Python code of PaddleOCR follows [PEP8 Specification]( https://www.python.org/dev/peps/pep-0008/ ), some of the key concerns include the following
- Space
- Spaces should be added after commas, semicolons, colons, not before them
```python
# true:
print(x, y)
# false:
print(x , y)
```
- When specifying a keyword parameter or default parameter value in a function, do not use spaces on both sides of it
```python
# true:
def complex(real, imag=0.0)
# false:
def complex(real, imag = 0.0)
```
- comment
- Inline comments: inline comments are indicated by the` # `sign. Two spaces should be left between code and` # `, and one space should be left between` # `and comments, for example
```python
x = x + 1 # Compensate for border
```
- Functions and methods: The definition of each function should include the following:
- Function description: Utility, input and output of function
- Args: Name and description of each parameter
- Returns: The meaning and type of the return value
Retrieves rows pertaining to the given keys from the Table instance
represented by big_table. Silly things may happen if
other_silly_variable is not None.
Args:
big_table: An open Bigtable Table instance.
keys: A sequence of strings representing the key of each table row
to fetch.
other_silly_variable: Another optional variable, that has a much
longer name than the other args, and which does nothing.
Returns:
A dict mapping keys to the corresponding table row data
fetched. Each row is represented as a tuple of strings. For
example:
{'Serak': ('Rigel VII', 'Preparer'),
'Zim': ('Irk', 'Invader'),
'Lrrr': ('Omicron Persei 8', 'Emperor')}
If a key from the keys argument is missing from the dictionary,
then that row was not found in the table.
"""
pass
```
<a name="Appendix2"></a>
## Appendix 2: Document Specification
### 2.1 Overall Description
- Document Location: If you add new features to your original Markdown file, please **Do not re-create** a new file. If you don't know where to add it, you can first PR the code and then ask the official in commit.
- New Markdown Document Name: Describe the content of the document in English, typically a combination of lowercase letters and underscores, such as `add_New_Algorithm.md`
- New Markdown Document Format: Catalog - Body - FAQ
> The directory generation method can use [this site](https://ecotrust-canada.github.io/markdown-toc/ ) Automatically extract directories after copying MD contents, and then add `<a name='XXXX'></a> before each heading of the MD file
- English and Chinese: Any changes or additions to the document need to be made in both Chinese and English documents.
### 2.2 Format Specification
- Title format: The document title format follows the format of: Arabic decimal point combination-space-title (for example, `2.1 XXXX`, `2.XXXX`)
- Code block: Displays code in code block format that needs to be run, describing the meaning of command parameters before the code block. for example:
> Pipeline of detection + direction Classify + recognition: Vertical text can be recognized after set direction classifier parameters`--use_angle_cls true`.
- Variable Rrferences: If code variables or command parameters are referenced in line, they need to be represented in line code, for example, above `--use_angle_cls true` with one space in front and one space in back
- Uniform naming: e.g. PP-OCRv2, PP-OCR mobile, `paddleocr` whl package, PPOCRLabel, Paddle Lite, etc.
- Supplementary notes: Supplementary notes by reference format `>`.
- Picture: If a picture is added to the description document, specify the naming of the picture (describing its content) and add the picture under `doc/`.
- Title: Capitalize the first letter of each word in the title.
<a name="Appendix3"></a>
## Appendix 3: Pull Request Description
### 3.1 PaddleOCR Branch Description
PaddleOCR will maintain two branches in the future, one for each:
- release/x.x family branch: stable release version branch, also the default branch. PaddleOCR releases a new release branch based on feature updates and adapts to the release version of Paddle. As versions iterate, more and more release/x.x family branches are maintained by default with the latest version of the release branch.
- dygraph branch: For the development branch, adapts the dygraph version of the Paddle dynamic graph to primarily develop new functionality. If you need to redevelop, choose the dygraph branch. To ensure that the dygraph branch pulls out the release/x.x branch when needed, the code for the dygraph branch can only use the valid API in the latest release branch of Paddle. That is, if a new API has been developed in the Paddle dygraph branch but has not yet appeared in the release branch code, do not use it in Paddle OCR. In addition, performance optimization, parameter tuning, policy updates that do not involve API can be developed normally.
The historical branch of PaddleOCR will no longer be maintained in the future. These branches will continue to be maintained, considering that some of you may still be using them:
- Develop branch: This branch was used for the development and testing of static diagrams and is currently compatible with version >=1.7. If you have special needs, you can also use this branch to accommodate older versions of Paddle, but you won't update your code until you fix the bug.
PaddleOCR welcomes you to actively contribute code to repo. Here are some basic processes for contributing code.
### 3.2 PaddleOCR Code Submission Process And Specification
> If you are familiar with Git use, you can jump directly to [Some Conventions For Submitting Code in 3.2.10](#Some_conventions_for_submitting_code)
#### 3.2.1 Create Your `Remote Repo`
- In PaddleOCR [GitHub Home]( https://github.com/PaddlePaddle/PaddleOCR ) Click the `Fork` button in the upper left corner to create a `remote repo`in your personal directory, such as ` https://github.com/ {your_name}/PaddleOCR`.
Only the information of the clone `remote repo`, i.e. the PaddleOCR under your username, is available. Due to the change in Github's login method, you need to reconfigure the `remote repo` address by means of a Token. The token is generated as follows:
1. Find Personal Access Tokens: Click on your avatar in the upper right corner of the Github page and choose Settings --> Developer settings --> Personal access tokens,
2. Click Generate new token: Fill in the token name in Note, such as 'paddle'. In Select scopes, select repo (required), admin:repo_hook, delete_repo, etc. You can check them according to your needs. Then click Generate token to generate the token, and finally copy the generated token.
Delete the original origin configuration
```
git remote rm origin
```
Change the remote branch to `https://oauth2:{token}@github.com/{your_name}/PaddleOCR.git`. For example, if the token value is 12345 and your user name is PPOCR, run the following command
This is mainly to keep the local repository up to date when subsequent pull request (PR) submissions are made.
#### 3.2.3 Create Local Branch
First get the latest code of upstream, then create a new_branch branch based on the dygraph of the upstream repo (upstream).
```
git fetch upstream
git checkout -b new_branch upstream/dygraph
```
> If for a newly forked PaddleOCR project, the user's remote repo (origin) has the same branch updates as the upstream repository (upstream), you can also create a new local branch based on the default branch of the origin repo or a specified branch with the following command
>
> ```
> # Create new_branch branch on user remote repo (origin) based on develop branch
> git checkout -b new_branch origin/develop
> # Create new_branch branch based on upstream remote repo develop branch
> # If you need to create a new branch from upstream,
> # you need to first use git fetch upstream to get upstream code
> git checkout -b new_branch upstream/develop
> ```
The final switch to the new branch is displayed with the following output information.
```
Branch new_branch set up to track remote branch develop from upstream.
Switched to a new branch 'new_branch'
```
After switching branches, file changes can be made on this branch
#### 3.2.4 Use Pre-Commit Hook
Paddle developers use the pre-commit tool to manage Git pre-submit hooks. It helps us format the source code (C++, Python) and automatically check for basic things (such as having only one EOL per file, not adding large files to Git) before committing it.
The pre-commit test is part of the unit test in Travis-CI. PR that does not satisfy the hook cannot be submitted to PaddleOCR. Install it first and run it in the current directory：
```
pip install pre-commit
pre-commit install
```
> 1. Paddle uses clang-format to adjust the C/C++ source code format. Make sure the `clang-format` version is above 3.8.
>
> 2. Yapf installed through pip install pre-commit is slightly different from conda install-c conda-forge pre-commit, and PaddleOCR developers use `pip install pre-commit`.
#### 3.2.5 Modify And Submit Code
If you make some changes on `README.Md ` on PaddleOCR, you can view the changed file through `git status`, and then add the changed file using `git add`。
```
git status # View change files
git add README.md
pre-commit
```
Repeat these steps until the pre-comit format check does not error. As shown below.
Use the following command to complete the submission.
```
git commit -m "your commit info"
```
#### 3.2.6 Keep Local Repo Up To Date
Get the latest code for upstream and update the current branch. Here the upstream comes from section 2.2, `Connecting to a remote repo`.
```
git fetch upstream
# If you want to commit to another branch, you need to pull code from another branch of upstream, here is develop
git pull upstream develop
```
#### 3.2.7 Push To Remote Repo
```
git push origin new_branch
```
#### 3.2.7 Submit Pull Request
Click the new pull request to select the local branch and the target branch, as shown in the following figure. In the description of PR, fill in the functions completed by the PR. Next, wait for review, and if you need to modify something, update the corresponding branch in origin with the steps above.
#### 3.2.8 Sign CLA Agreement And Pass Unit Tests
- Signing the CLA When submitting a Pull Request to PaddlePaddle for the first time, you need to sign a CLA (Contributor License Agreement) agreement to ensure that your code can be incorporated as follows:
1. Please check the Check section in PR, find the license/cla, and click on the right detail to enter the CLA website
2. Click Sign in with GitHub to agree on the CLA website and when clicked, it will jump back to your Pull Request page
#### 3.2.9 Delete Branch
- Remove remote branch
After PR is merged into the main repo, we can delete the branch of the remote repofrom the PR page.
You can also use `git push origin:branch name` to delete remote branches, such as:
```
git push origin :new_branch
```
- Delete local branch
```
# Switch to the development branch, otherwise the current branch cannot be deleted
In order for official maintainers to better focus on the code itself when reviewing it, please follow the following conventions each time you submit your code:
1）Please ensure that the unit tests in Travis-CI pass smoothly. If not, indicate that there is a problem with the submitted code, and the official maintainer generally does not review it.
2）Before submitting a Pull Request.
- Note the number of commits.
Reason: If you only modify one file and submit more than a dozen commits, each commit will only make a few modifications, which can be very confusing to the reviewer. The reviewer needs to look at each commit individually to see what changes have been made, and does not exclude the fact that changes between commits overlap each other.
Suggestion: Keep as few commits as possible each time you submit, and supplement your last commit with git commit --amend. For multiple commits that have been Push to a remote warehouse, you can refer to [squash commits after push](https://stackoverflow.com/questions/5667884/how-to-squash-commits-in-git-after-they-have-been-pushed ).
- Note the name of each commit: it should reflect the content of the current commit, not be too arbitrary.
3） If you have solved a problem, add in the first comment box of the Pull Request:fix #issue_number，This will automatically close the corresponding Issue when the Pull Request is merged. Key words include:close, closes, closed, fix, fixes, fixed, resolve, resolves, resolved,please choose the right vocabulary. Detailed reference [Closing issues via commit messages](https://help.github.com/articles/closing-issues-via-commit-messages).
In addition, in response to the reviewer's comments, you are requested to abide by the following conventions:
1） Each review comment from an official maintainer would like a response, which would better enhance the contribution of the open source community.
- If you agree to the review opinion and modify it accordingly, give a simple Done.
- If you disagree with the review, please give your own reasons for refuting.
2）If there are many reviews:
- Please give an overview of the changes.
- Please reply with `start a review', not directly. The reason is that each reply sends an e-mail message, which can cause a mail disaster.
Thank you for your support and interest in PaddleOCR. The goal of PaddleOCR is to build a professional, harmonious and supportive open source community with developers. This document presents existing community contributions, explanations for various contributions, and new opportunities and processes to make the contribution process more efficient and clear.
PaddleOCR wants to help any developer with a dream realize their vision and enjoy the joy of creating value through the power of AI.
-[scr2txt](https://github.com/lstwzd/scr2txt)：Screenshot to Text tool (@ [lstwzd](https://github.com/lstwzd))
-[AI Studio project](https://aistudio.baidu.com/aistudio/projectdetail/1054614?channelType=0&channel=0)：English video automatically generates subtitles( @ [叶月水狐](https://aistudio.baidu.com/aistudio/personalcenter/thirdview/322052))
#### 1.1.2 Vertical Scene Tools
-[id_card_ocr](https://github.com/baseli/id_card_ocr)：Identification of copy of ID card(@ [baseli](https://github.com/baseli))
-[Paddle_Table_Image_Reader](https://github.com/thunder95/Paddle_Table_Image_Reader): A data assistant that can read tables and pictures(@ [thunder95](https://github.com/thunder95]))
#### 1.1.3 Pre And Post Processing
-[paddleOCRCorrectOutputs](https://github.com/yuranusduke/paddleOCRCorrectOutputs)：Get the key-value of OCR recognition result (@ [yuranusduke](https://github.com/yuranusduke))
### 1.2 New Features For PaddleOCR
- Thanks [authorfu](https://github.com/authorfu) for contributing Android([#340](https://github.com/PaddlePaddle/PaddleOCR/pull/340)) and [xiadeye](https://github.com/xiadeye) for contributing IOS demo code([#325](https://github.com/PaddlePaddle/PaddleOCR/pull/325)).
- Thanks [tangmq](https://gitee.com/tangmq) for adding docker deployment service to PaddleOCR to support quick release of callable restful API services([#507](https://github.com/PaddlePaddle/PaddleOCR/pull/507)).
- Thanks [lijinhan](https://github.com/lijinhan) for adding Java springboot to PaddleOCR and call OCR hubserving interface to complete the use of OCR service deployment([#1027](https://github.com/PaddlePaddle/PaddleOCR/pull/1027)).
- Thanks [Evezerest](https://github.com/Evezerest), [ninetailskim](https://github.com/ninetailskim), [edencfc](https://github.com/edencfc), [BeyondYourself](https://github.com/BeyondYourself), [1084667371](https://github.com/1084667371) for contributing complete code of [PPOCRLabel](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/PPOCRLabel/README_ch.md).
### 1.3 Code And Document Optimization
- Thanks [zhangxin](https://github.com/ZhangXinNan)([Blog](https://blog.csdn.net/sdlypyzq)) for contributing new visualization methods and adding .gitgnore, handling the problem of manually setting the PYTHONPATH environment variable([#210](https://github.com/PaddlePaddle/PaddleOCR/pull/210)).
- Thanks [lyl120117](https://github.com/lyl120117) for contributing code to print network structure([#304](https://github.com/PaddlePaddle/PaddleOCR/pull/304)).
- Thanks [BeyondYourself](https://github.com/BeyondYourself) for making a lot of great suggestions for PaddleOCR and simplifying some code styles of paddleocr([so many commits)](https://github.com/PaddlePaddle/PaddleOCR/commits?author=BeyondYourself).
- Thanks [Khanh Tran](https://github.com/xxxpsyduck) and [Karl Horky](https://github.com/karlhorky) for contributing modifing English documents.
### 1.4 Multilingual Corpus
- Thanks [xiangyubo](https://github.com/xiangyubo) for contributing handwritting Chinese OCR dataset([#321](https://github.com/PaddlePaddle/PaddleOCR/pull/321)).
- Thanks [Mejans](https://github.com/Mejans) for contributing dictionary and corpus of the new language Occitan to PaddleOCR([#954](https://github.com/PaddlePaddle/PaddleOCR/pull/954)).
## 2. Contribution Illustrating
### 2.1 New Function Class
PaddleOCR welcomes community contributions to various services, deployment examples and software applications with paddleOCR as the core. Certified community contributions will be added to the above community contribution table to increase exposure for the majority of developers, which is also the glory of PaddleOCR, including:
- Project form: the project code certified by the official community shall have good specifications and structure, and shall be equipped with a detailed README.md, which describes how to use the project. Through add a line 'paddleocr' to the requirements.txt, which can be automatically included in the usedby of paddleocr.
- Integration method: if it is an update to the existing PaddleOCR tool, it will be integrated into the main repo. If a new function is expanded for paddleocr, please contact the official personnel first to confirm whether the project is integrated into the master repo, *even if the new function is not integrated into the master repo, we will also increase the exposure of your personal project in the way of community contribution.*
### 2.2 Code Optimization
If you encounter code bugs and unexpected functions when using PaddleOCR, you can contribute your modifications to PaddleOCR, including:
- Python code specifications are available for reference [Appendix 1:Python code specifications](./code_and_doc_en.md/#Appendix1).
- Before submitting the code, please confirm again and again that no new bugs will be introduced, and describe the optimization points in the PR. If the PR solves an issue, please connect to the issue in the PR. All PR shall comply with the requirements in Appendix [3.2.10 Some conventions for submitting code.](./code_and_doc_en.md/#Some conventions for submitting code)
- Please refer to the below before submitting. If you are not familiar with the git submission process, you can also refer to Section 3.2 of [Appendix 3: description of Pull Request](./code_and_doc_en.md/#Appendix3).If you are not familiar with the git submission process, you can also refer to Section 3.2 of Appendix 3.
**Finally, please add the label Third Party in the title of PR and @ Everest in the description , PR with this label will be treated with high priority`[third-part]`.**
### 2.3 Document Optimization
If you encounter problems such as unclear document description, missing description and invalid link when using PaddleOCR, you can contribute your modifications to PaddleOCR. For document writing specifications, please refer to [Appendix 2: document specifications](./code_and_doc_en.md/#Appendix2). **Finally, please add the label Third Party in the title of PR and @ Everest in the description , PR with this label will be treated with high priority`[third-party].**
## 3. More Contribution Opportunities
We encourage developers to use PaddleOCR to realize their ideas. At the same time, we also list some valuable development directions after analysis, which are collected in the regular season of community projects as a whole.
## 4. Contact Us
We very much welcome developers to contact us before they intend to contribute code, documents, corpus and other contents to PaddleOCR, which can greatly reduce the communication cost in the PR process. At the same time, if you find some ideas difficult to realize personally, we can also recruit like-minded developers for the project in the form of SIG. Projects funded through SIG channels will receive deep R & D support and operational resources (such as official account publicity, live broadcast lessons, etc.).
Our recommended contribution process is:
- By adding the `[Third Party]` mark in the topic of GitHub issue, explain the problems encountered (and the ideas to solve) or the functions to be expanded, and wait for the reply of the person on duty. For example, ` [Third Party] contributes IOS examples to PaddleOCR`.
- After communicating with us and confirming that the technical scheme or bugs and optimization points are correct, add functions or modify them accordingly, and the codes and documents shall comply with relevant specifications.
- PR links to the above issue and waits for review.
## 5. Thanks And Follow-Up
- After the code is combined, the information will be updated in the first section of this document. The default link is GitHub name and home page. If you need to change the home page, you can also contact us.
- New important function classes will be advertised in the user group and enjoy the honor of the open source community.
-**If you have a PaddleOCR based project that does not appear in the above list, follow `4. Contact Us` .**
In OCR recognition, CRNN is a text recognition algorithm widely applied in the industry. In the training phase, it uses CTCLoss to calculate the network loss. In the inference phase, it uses CTCDecode to obtain the decoding result. Although the CRNN algorithm has been proven to achieve reliable recognition results in actual business, users have endless requirements for recognition accuracy. So how to improve the accuracy of text recognition? Taking CTCLoss as the starting point, this paper explores the improved fusion scheme of CTCLoss from three different perspectives: Hard Example Mining, Multi-task Learning, and Metric Learning. Based on the exploration, we propose EnhancedCTCLoss, which includes the following 3 components: Focal-CTC Loss, A-CTC Loss, C-CTC Loss.
## 1. Focal-CTC Loss
Focal Loss was proposed by the paper, "[Focal Loss for Dense Object Detection](https://arxiv.org/abs/1708.02002)". When the loss was first proposed, it was mainly to solve the problem of a serious imbalance in the ratio of positive and negative samples in one-stage target detection. This loss function reduces the weight of a large number of simple negative samples in training and also can be understood as a kind of difficult sample mining.
The form of the loss function is as follows:
<divalign="center">
<imgsrc="./focal_loss_formula.png"width = "600"/>
</div>
Among them, y' is the output of the activation function, and the value is between 0-1. It adds a modulation factor (1-y’)^γ and a balance factor α on the basis of the original cross-entropy loss. When α = 1, y = 1, the comparison between the loss function and the cross-entropy loss is shown in the following figure:
<divalign="center">
<imgsrc="./focal_loss_image.png"width = "600"/>
</div>
As can be seen from the above figure, when γ > 0, the adjustment coefficient (1-y’)^γ gives smaller weight to the easy-to-classify sample loss, making the network pay more attention to the difficult and misclassified samples. The adjustment factor γ is used to adjust the rate at which the weight of simple samples decreases. When γ = 0, it is the cross-entropy loss function. When γ increases, the influence of the adjustment factor will also increase. Experiments revealed that 2 is the optimal value of γ. The balance factor α is used to balance the uneven proportions of the positive and negative samples. In the text, α is taken as 0.25.
For the classic CTC algorithm, suppose a certain feature sequence (f<sub>1</sub>, f<sub>2</sub>, ......f<sub>t</sub>), after CTC decoding, the probability that the result is equal to label is y', then the probability that the CTC decoding result is not equal to label is (1-y'); it is not difficult to find that the CTCLoss value and y' have the following relationship:
<divalign="center">
<imgsrc="./equation_ctcloss.png"width = "250"/>
</div>
Combining the idea of Focal Loss, assigning larger weights to difficult samples and smaller weights to simple samples can make the network focus more on the mining of difficult samples and further improve the accuracy of recognition. Therefore, we propose Focal-CTC Loss. Its definition is as follows:
<divalign="center">
<imgsrc="./equation_focal_ctc.png"width = "500"/>
</div>
In the experiment, the value of γ is 2, α = 1, see this for specific implementation: [rec_ctc_loss.py](../../ppocr/losses/rec_ctc_loss.py)
## 2. A-CTC Loss
A-CTC Loss is short for CTC Loss + ACE Loss. Among them, ACE Loss was proposed by the paper, “[Aggregation Cross-Entropy for Sequence Recognition](https://arxiv.org/abs/1904.08364)”. Compared with CTCLoss, ACE Loss has the following two advantages:
+ ACE Loss can solve the recognition problem of 2-D text, while CTCLoss can only process 1-D text
+ ACE Loss is better than CTC loss in time complexity and space complexity
The advantages and disadvantages of the OCR recognition algorithm summarized by the predecessors are shown in the following figure:
<divalign="center">
<imgsrc="./rec_algo_compare.png"width = "1000"/>
</div>
Although ACELoss does handle 2D predictions, as shown in the figure above, and has advantages in memory usage and inference speed, in practice, we found that using ACELoss alone, the recognition effect is not as good as CTCLoss. Consequently, we tried to combine CTCLoss and ACELoss, and CTCLoss is the mainstay while ACELoss acts as an auxiliary supervision loss. This attempt has achieved better results. On our internal experimental data set, compared to using CTCLoss alone, the recognition accuracy can be improved by about 1%.
A_CTC Loss is defined as follows:
<divalign="center">
<imgsrc="./equation_a_ctc.png"width = "300"/>
</div>
In the experiment, λ = 0.1. See the ACE loss implementation code: [ace_loss.py](../../ppocr/losses/ace_loss.py)
## 3. C-CTC Loss
C-CTC Loss is short for CTC Loss + Center Loss. Among them, Center Loss was proposed by the paper, “[A Discriminative Feature Learning Approach for Deep Face Recognition](https://link.springer.com/chapter/10.1007/978-3-319-46478-7_31)“. It was first used in face recognition tasks to increase the distance between classes and reduce the distance within classes. It is an earlier and also widely used algorithm.
In the task of Chinese OCR recognition, through the analysis of bad cases, we found that a major difficulty in Chinese recognition is that there are many similar characters, which are easy to misunderstand. From this, we thought about whether we can learn from the idea of n to increase the class spacing of similar characters, to improve recognition accuracy. However, Metric Learning is mainly used in the field of image recognition, and the label of the training data is a fixed value; for OCR recognition, it is a sequence recognition task essentially, and there is no explicit alignment between features and labels. Therefore, how to combine the two is still a direction worth exploring.
By trying Arcmargin, Cosmargin and other methods, we finally found that Centerloss can help further improve the accuracy of recognition. C_CTC Loss is defined as follows:
<divalign="center">
<imgsrc="./equation_c_ctc.png"width = "300"/>
</div>
In the experiment, we set λ=0.25. See the center_loss implementation code: [center_loss.py](../../ppocr/losses/center_loss.py)
It is worth mentioning that in C-CTC Loss, choosing to initialize the Center randomly does not bring significant improvement. Our Center initialization method is as follows:
+ Based on the original CTCLoss, a network N is obtained by training
+ Select the training set, identify the completely correct part, and form the set G
+ Send each sample in G to the network, perform forward calculation, and extract the correspondence between the input of the last FC layer (ie feature) and the result of argmax calculation (ie index)
+ Aggregate features with the same index, calculate the average, and get the initial center of each character.
Taking the configuration file `configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml` as an example, the center extraction command is as follows:
After running, `train_center.pkl` will be generated in the main directory of PaddleOCR.
## 4. Experiment
For the above three solutions, we conducted training and evaluation based on Baidu's internal data set. The experimental conditions are shown in the following table:
| algorithm | Focal_CTC | A_CTC | C-CTC |
| :-------- | :-------- | ----: | :---: |
| gain | +0.3% | +0.7% | +1.7% |
Based on the above experimental conclusions, we adopted the C-CTC strategy in PP-OCRv2. It is worth mentioning that, because PP-OCRv2 deals with the recognition task of 6625 Chinese characters, the character set is relatively large and there are many similar characters, so the C-CTC solution brings a significant improvement on this task. But if you switch to other OCR recognition tasks, the conclusion may be different. You can try Focal-CTC, A-CTC, C-CTC, and the combined solution EnhancedCTC. We believe it will bring different degrees of improvement.
The unified combined plan is shown in the following file: [rec_enhanced_ctc_loss.py](../../ppocr/losses/rec_enhanced_ctc_loss.py)
|ch_PP-OCRv2_rec_slim|[New] Slim qunatization with distillation lightweight model, supporting Chinese, English, multilingual text detection|[ch_PP-OCRv2_rec.yml](../../configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml)| 9M |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_train.tar) |
|ch_PP-OCRv2_rec_slim|[New] Slim qunatization with distillation lightweight model, supporting Chinese, English, multilingual text recognition|[ch_PP-OCRv2_rec.yml](../../configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml)| 9M |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_train.tar) |
|ch_PP-OCRv2_rec|[New] Original lightweight model, supporting Chinese, English, multilingual text detection|[ch_PP-OCRv2_rec.yml](../../configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml)|8.5M|[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_train.tar) |
|ch_PP-OCRv2_rec|[New] Original lightweight model, supporting Chinese, English, multilingual text recognition|[ch_PP-OCRv2_rec_distillation.yml](../../configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec_distillation.yml)|8.5M|[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_train.tar) |
|ch_ppocr_mobile_slim_v2.0_rec|Slim pruned and quantized lightweight model, supporting Chinese, English and number recognition|[rec_chinese_lite_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml)| 6M | [inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_train.tar) |
|ch_ppocr_mobile_slim_v2.0_rec|Slim pruned and quantized lightweight model, supporting Chinese, English and number recognition|[rec_chinese_lite_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml)| 6M | [inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_train.tar) |
|ch_ppocr_mobile_v2.0_rec|Original lightweight model, supporting Chinese, English and number recognition|[rec_chinese_lite_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml)|5.2M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_train.tar) / [pre-trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_pre.tar) |
|ch_ppocr_mobile_v2.0_rec|Original lightweight model, supporting Chinese, English and number recognition|[rec_chinese_lite_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml)|5.2M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_train.tar) / [pre-trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_pre.tar) |
|ch_ppocr_server_v2.0_rec|General model, supporting Chinese, English and number recognition|[rec_chinese_common_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_common_train_v2.0.yml)|94.8M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_train.tar) / [pre-trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_pre.tar) |
|ch_ppocr_server_v2.0_rec|General model, supporting Chinese, English and number recognition|[rec_chinese_common_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_common_train_v2.0.yml)|94.8M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_train.tar) / [pre-trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_pre.tar) |
This section provides a tutorial example on how to quickly use, train, and evaluate a key information extraction(KIE) model, [SDMGR](https://arxiv.org/abs/2103.14470), in PaddleOCR.
[SDMGR(Spatial Dual-Modality Graph Reasoning)](https://arxiv.org/abs/2103.14470) is a KIE algorithm that classifies each detected text region into predefined categories, such as order ID, invoice number, amount, and etc.
*[1. Quick Use](#1-----)
*[2. Model Training](#2-----)
*[3. Model Evaluation](#3-----)
<aname="1-----"></a>
## 1. Quick Use
[Wildreceipt dataset](https://paperswithcode.com/dataset/wildreceipt) is used for this tutorial. It contains 1765 photos, with 25 classes, and 50000 text boxes, which can be downloaded by wget:
The prediction result is saved as `./output/sdmgr_kie/predicts_kie.txt`, and the visualization results are saved in the folder`/output/sdmgr_kie/kie_results/`.
The visualization results are shown in the figure below:
<divalign="center">
<imgsrc="./imgs/0.png"width="800">
</div>
<aname="2-----"></a>
## 2. Model Training
Create a softlink to the folder, `PaddleOCR/train_data`:
```shell
cd PaddleOCR/ &&mkdir train_data &&cd train_data
ln-s ../../wildreceipt ./
```
The configuration file used for training is `configs/kie/kie_unet_sdmgr.yml`. The default training data path in the configuration file is `train_data/wildreceipt`. After preparing the data, you can execute the model training with the following command:
Note: The above model is trained on the PubLayNet dataset and only supports English scanning scenarios. If you need to identify other scenarios, you need to train the model yourself and replace the three fields `det_model_dir`, `rec_model_dir`, `table_model_dir`.
Note: The above model is trained on the PubLayNet dataset and only supports English scanning scenarios. If you need to identify other scenarios, you need to train the model yourself and replace the three fields `det_model_dir`, `rec_model_dir`, `table_model_dir`.
After running, the excel sheet of each picture will be saved in the directory specified by the output field
After running, the excel sheet of each picture will be saved in the directory specified by the output field
...
@@ -52,10 +55,13 @@ After running, the excel sheet of each picture will be saved in the directory sp
...
@@ -52,10 +55,13 @@ After running, the excel sheet of each picture will be saved in the directory sp
In this chapter, we only introduce the training of the table structure model, For model training of [text detection](../../doc/doc_en/detection_en.md) and [text recognition](../../doc/doc_en/recognition_en.md), please refer to the corresponding documents
In this chapter, we only introduce the training of the table structure model, For model training of [text detection](../../doc/doc_en/detection_en.md) and [text recognition](../../doc/doc_en/recognition_en.md), please refer to the corresponding documents
#### data preparation
#### data preparation
The training data uses public data set [PubTabNet](https://arxiv.org/abs/1911.10683), Can be downloaded from the official [website](https://github.com/ibm-aur-nlp/PubTabNet) 。The PubTabNet data set contains about 500,000 images, as well as annotations in html format。
The training data uses public data set [PubTabNet](https://arxiv.org/abs/1911.10683), Can be downloaded from the official [website](https://github.com/ibm-aur-nlp/PubTabNet) 。The PubTabNet data set contains about 500,000 images, as well as annotations in html format。
#### Start training
#### Start training
*If you are installing the cpu version of paddle, please modify the `use_gpu` field in the configuration file to false*
*If you are installing the cpu version of paddle, please modify the `use_gpu` field in the configuration file to false*
The table uses [TEDS(Tree-Edit-Distance-based Similarity)](https://github.com/ibm-aur-nlp/PubTabNet/tree/master/src) as the evaluation metric of the model. Before the model evaluation, the three models in the pipeline need to be exported as inference models (we have provided them), and the gt for evaluation needs to be prepared. Examples of gt are as follows:
The table uses [TEDS(Tree-Edit-Distance-based Similarity)](https://github.com/ibm-aur-nlp/PubTabNet/tree/master/src) as the evaluation metric of the model. Before the model evaluation, the three models in the pipeline need to be exported as inference models (we have provided them), and the gt for evaluation needs to be prepared. Examples of gt are as follows:
Document Visual Q&A, mainly for the image content of the question and answer, DOC-VQA is a type of VQA task, DOC-VQA mainly asks questions about the textual content of text images.
The DOC-VQA algorithm in PP-Structure is developed based on PaddleNLP natural language processing algorithm library.
The main features are as follows:
- Integrated LayoutXLM model and PP-OCR prediction engine.
- Support Semantic Entity Recognition (SER) and Relation Extraction (RE) tasks based on multi-modal methods. Based on SER task, text recognition and classification in images can be completed. Based on THE RE task, we can extract the relation of the text content in the image, such as judge the problem pair.
- Support custom training for SER and RE tasks.
- Support OCR+SER end-to-end system prediction and evaluation.
- Support OCR+SER+RE end-to-end system prediction.
**Note**: This project is based on the open source implementation of [LayoutXLM](https://arxiv.org/pdf/2104.08836.pdf) on Paddle 2.2, and at the same time, after in-depth polishing by the flying Paddle team and the Industrial and **Commercial Bank of China** in the scene of real estate certificate, jointly open source.
## 1.Performance
We evaluated the algorithm on [XFUN](https://github.com/doc-analysis/XFUND) 's Chinese data set, and the performance is as follows
| Model | Task | F1 | Model Download Link |
|:---:|:---:|:---:| :---:|
| LayoutXLM | RE | 0.7113 | [Link](https://paddleocr.bj.bcebos.com/pplayout/PP-Layout_v1.0_re_pretrained.tar) |
| LayoutXLM | SER | 0.9056 | [Link](https://paddleocr.bj.bcebos.com/pplayout/PP-Layout_v1.0_ser_pretrained.tar) |
| LayoutLM | SER | 0.78 | [Link](https://paddleocr.bj.bcebos.com/pplayout/LayoutLM_ser_pretrained.tar) |
## 2.Demonstration
**Note**: the test images are from the xfun dataset.
In the figure, the red box represents the question, the blue box represents the answer, and the question and answer are connected by green lines. The corresponding category and OCR recognition results are also marked at the top left of the OCR detection box.
# Note: the code cloud hosting code may not be able to synchronize the update of this GitHub project in real time, with a delay of 3 ~ 5 days. Please give priority to the recommended method.
```
-**(3) Install PaddleNLP**
```bash
# You need to use the latest code version of paddlenlp for installation
Download address of processed xfun Chinese dataset: [https://paddleocr.bj.bcebos.com/dataset/XFUND.tar](https://paddleocr.bj.bcebos.com/dataset/XFUND.tar)。
Download and unzip the dataset, and then place the dataset in the current directory.
If you want to convert data sets in other languages in xfun, you can refer to [xfun data conversion script.](helper/trans_xfun_data.py))
If you want to experience the prediction process directly, you can download the pre training model provided by us, skip the training process and predict directly.
It will end up in output_res The visual image of the prediction result and the text file of the prediction result are saved in the res directory. The file name is infer_ results.txt.
The visual image of the prediction result and the text file of the prediction result are saved in the output_res file folder, the file name is`infer_results.txt`。
* Concatenation results using OCR engine + SER+ RE
