Scanned Document Rotation Correction

The project creates the models and service API for predicting scanned document images' angles ranging between -90° (counter clockwise) to 90° (clockwise) from the vertical.

Setup Environment For Model Training and Testing

conda env create -f conda_env.yml -p <path of env>

conda activate <env_name> 

pip install -r requirements.txt

Training

1. Dataset preparation

   • Convert PDF documents to images.

   • Download scanned document images from website.

   • Etc.

2. Data augmentation (Coming soon)

    python src/data_augmentation.py
   

3. Save dataset to Torch format (.pt) (Coming soon)

    python src/create_dataset.py
   

4. Training

   4.1 Edit dataset path

   4.2 Training

    python src/training.py
   

   4.3 TensorBoard Running

    tensorboard --logdir=<log dir> --port <port>
    
    tensorboard --logdir=runs --port 8000
   

Models

Download rotation_net.onnx into models/

Running Service API with docker

1. Start up application by running.

    docker-compose up
   

   Or

    docker-compose -f .\docker-compose.yml up
   

2. Open http://127.0.0.1:9000/docs API documentation.

3. Send POST Request to http://127.0.0.1:9000/angle or api_endpoint/angle with Json format.

    {
      "image" : "image_base_64"
    } 
   

4. Response is angle in range from -90° to 90° (counter clockwise to clockwise).

    {
        "angle": 0.7157974243164062
    }
   

Testing Output

Success Cases	Failed Cases

Todo list

• Add data_augmentation script

• Add create dataset script

• Impove model accuracy

Contributing

Anyone's participation is welcome! Open an issue or submit PRs.