This is a simple image classifier based on the Inception model to get predictions on whether a supplied image is a document or not.
Install Python dependencies using PIP:
pip install -r requirements.txt
Add training images in the training/images directory like shown below. Due to the usage of the Inception model you can achieve great results with a relatively small data set (~100 images for each category).
├── training
│ ├── images
│ │ ├── documents [your training images]
│ │ └── random [your training images]
Once we have the training images we can start the process of retraining the Inception model.
scripts/training.sh
Notice that this image classifier currently only works with JPEG images.
$ python src/prediction.py <YOUR_IMAGE_URL>
> document (score = 0.99978)
> random (score = 0.00022)
For development purposes you can run a simple REST endpoint to serve predictions. For serious production use something like TensorFlow Serving is highly recommended.
export FLASK_APP="api.py"
export FLASK_DEBUG=1
flask run