Responsive Doc. scanner using U^2-Net, Textcleaner and Tesseract

Toolset

• U^2-Net is used for background removal
• Textcleaner is used for image cleaning and line deskew (max 5 degrees)
• Tesseract is used for text angle rotation
• Deskew is used for line deskew (between 5 and 45 degrees)

Examples

Tested one document on smartphone camera with different angles

To build & deploy

1. Clone thee repo
2. Download the model: check app/saved_models/README.md
3. Build Docker image : docker build -t <REPOSITORY-NAME>/<IMAGE>:<TAG> .
4. Test locally : Run Docker image and check if api is working by running http://localhost:10000
   • CPU : docker run -it -v $PWD:/LOCAL/ -p 10000:80 <REPOSITORY-NAME>/<IMAGE>:<TAG>
   • GPU : docker run -it --gpus all -v $PWD:/LOCAL/ -p 10000:80 <REPOSITORY-NAME>/<IMAGE>:<TAG>
5. Push docker image to Dockerhub (optional):
   • Check: https://docs.docker.com/docker-hub/repos/ for account setup
   • Create in Dockerhub Repo similar to the name of yout Image ID : <REPOSITORY-NAME>
   • Run docker push <REPOSITORY-NAME>/<IMAGE>:<TAG>
6. Deploy to Cloud Run (optional):
   • Create your google cloud account
   • Push Docker Image to Google Container Registry
     • create new project called [PROJECT-ID]
     • Open Cloud shell in your Google account and run: docker pull <REPOSITORY-NAME>/<IMAGE>:<TAG> docker tag [IMAGE] gcr.io/[PROJECT-ID]/[IMAGE] docker push gcr.io/[PROJECT-ID]/[IMAGE] more detail in this link
   • Create CloudRun Service, and select Container that was created
     • Screenshot of the config - for demo purpose, it will be cost free
     
   • Click Deploy, and test the Api Url that will display

Limits and Areas for improvements

• Speed: It takes 7 to 10 seconds to process one image (serverless Cloud Run) With Gpu we can save 2 to 3 seconds (U^2-Net is 3 times faster)
• Textcleaner is slow(speed) but works good on image cleaning, but needs some manual fine-tuning a faster alternative can be used (Ex. Opencv)
• Taking pictures from angled positions is not supported, perspective transformation can be used but may deteriorate text quality
• U^2-Net limitations : the document should be centered in a contrasting color background (white background will not work)

References

• U^2-Net https://github.com/xuebinqin/U-2-Net.git
• Textcleaner http://www.fmwconcepts.com/imagemagick/textcleaner/
• Tesseract https://github.com/tesseract-ocr/tesseract
• Deskew https://github.com/sbrunner/deskew.git