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

Dataset preparation
- Convert PDF documents to images.
- Download scanned document images from website.
- Etc.
Data augmentation (Coming soon)
```
 python src/data_augmentation.py
```
Save dataset to Torch format (.pt) (Coming soon)
```
 python src/create_dataset.py
```

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

Start up application by running.

 docker-compose up

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

Open http://127.0.0.1:9000/docs API documentation.
Send POST Request to http://127.0.0.1:9000/angle or api_endpoint/angle with Json format.
```
 {
   "image" : "image_base_64"
 } 
```
Response is angle in range from -90° to 90° (counter clockwise to clockwise).
```
 {
     "angle": 0.7157974243164062
 }
```