pengjunjie/ragflow
1
0
Fork 1
mirror of https://github.com/infiniflow/ragflow.git synced 2025-12-08 12:32:30 +08:00
Code Issues Packages Projects Releases Wiki Activity

add ocr and recognizer demo, update README (#74)

Browse Source
This commit is contained in:
KevinHuSh
2024-02-26 19:51:35 +08:00
committed by GitHub
parent d1417102b6
commit d1c600d5d3
9 changed files with 525 additions and 73 deletions
Download Patch File Download Diff File Expand all files Collapse all files

62
deepdoc/README.md
View File

@ -11,7 +11,36 @@ English | [简体中文](./README_zh.md)
With a bunch of documents from various domains with various formats and along with diverse retrieval requirements,
an accurate analysis becomes a very challenge task. *Deep*Doc is born for that purpose.
There 2 parts in *Deep*Doc so far: vision and parser.
There are 2 parts in *Deep*Doc so far: vision and parser.
You can run the flowing test programs if you're interested in our results of OCR, layout recognition and TSR.
```bash
python deepdoc/vision/t_ocr.py -h
usage: t_ocr.py [-h] --inputs INPUTS [--output_dir OUTPUT_DIR]
options:
-h, --help show this help message and exit
--inputs INPUTS Directory where to store images or PDFs, or a file path to a single image or PDF
--output_dir OUTPUT_DIR
Directory where to store the output images. Default: './ocr_outputs'
```
```bash
python deepdoc/vision/t_recognizer.py -h
usage: t_recognizer.py [-h] --inputs INPUTS [--output_dir OUTPUT_DIR] [--threshold THRESHOLD] [--mode {layout,tsr}]
options:
-h, --help show this help message and exit
--inputs INPUTS Directory where to store images or PDFs, or a file path to a single image or PDF
--output_dir OUTPUT_DIR
Directory where to store the output images. Default: './layouts_outputs'
--threshold THRESHOLD
A threshold to filter out detections. Default: 0.5
--mode {layout,tsr} Task mode: layout recognition or table structure recognition
```
Our models are served on HuggingFace. If you have trouble downloading HuggingFace models, this might help!!
```bash
export HF_ENDPOINT=https://hf-mirror.com
```
<a name="2"></a>
## 2. Vision
@ -19,9 +48,14 @@ There 2 parts in *Deep*Doc so far: vision and parser.
We use vision information to resolve problems as human being.
- OCR. Since a lot of documents presented as images or at least be able to transform to image,
OCR is a very essential and fundamental or even universal solution for text extraction.
```bash
python deepdoc/vision/t_ocr.py --inputs=path_to_images_or_pdfs --output_dir=path_to_store_result
```
The inputs could be directory to images or PDF, or a image or PDF.
You can look into the folder 'path_to_store_result' where has images which demonstrate the positions of results,
txt files which contain the OCR text.
<div align="center" style="margin-top:20px;margin-bottom:20px;">
<img src="https://lh6.googleusercontent.com/2xdiSjaGWkZ71YdORc71Ujf7jCHmO6G-6ONklzGiUYEh3QZpjPo6MQ9eqEFX20am_cdW4Ck0YRraXEetXWnM08kJd99yhik13Cy0_YKUAq2zVGR15LzkovRAmK9iT4o3hcJ8dTpspaJKUwt6R4gN7So" width="300"/>
<img src="https://github.com/infiniflow/ragflow/assets/12318111/f25bee3d-aaf7-4102-baf5-d5208361d110" width="900"/>
</div>
- Layout recognition. Documents from different domain may have various layouts,
@ -39,11 +73,18 @@ We use vision information to resolve problems as human being.
- Footer
- Reference
- Equation
Have a try on the following command to see the layout detection results.
```bash
python deepdoc/vision/t_recognizer.py --inputs=path_to_images_or_pdfs --threshold=0.2 --mode=layout --output_dir=path_to_store_result
```
The inputs could be directory to images or PDF, or a image or PDF.
You can look into the folder 'path_to_store_result' where has images which demonstrate the detection results as following:
<div align="center" style="margin-top:20px;margin-bottom:20px;">
<img src="https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.7/ppstructure/docs/layout/layout.png?raw=true" width="900"/>
<img src="https://github.com/infiniflow/ragflow/assets/12318111/07e0f625-9b28-43d0-9fbb-5bf586cd286f" width="1000"/>
</div>
- Table Structure Recognition(TSR). Data table is a frequently used structure present data including numbers or text.
- Table Structure Recognition(TSR). Data table is a frequently used structure to present data including numbers or text.
And the structure of a table might be very complex, like hierarchy headers, spanning cells and projected row headers.
Along with TSR, we also reassemble the content into sentences which could be well comprehended by LLM.
We have five labels for TSR task:
@ -52,8 +93,15 @@ We use vision information to resolve problems as human being.
- Column header
- Projected row header
- Spanning cell
Have a try on the following command to see the layout detection results.
```bash
python deepdoc/vision/t_recognizer.py --inputs=path_to_images_or_pdfs --threshold=0.2 --mode=tsr --output_dir=path_to_store_result
```
The inputs could be directory to images or PDF, or a image or PDF.
You can look into the folder 'path_to_store_result' where has both images and html pages which demonstrate the detection results as following:
<div align="center" style="margin-top:20px;margin-bottom:20px;">
<img src="https://user-images.githubusercontent.com/10793386/139559159-cd23c972-8731-48ed-91df-f3f27e9f4d79.jpg" width="900"/>
<img src="https://github.com/infiniflow/ragflow/assets/12318111/cb24e81b-f2ba-49f3-ac09-883d75606f4c" width="1000"/>
</div>
<a name="3"></a>
@ -71,4 +119,4 @@ The résumé is a very complicated kind of document. A résumé which is compose
with various layouts could be resolved into structured data composed of nearly a hundred of fields.
We haven't opened the parser yet, as we open the processing method after parsing procedure.