Antialias with Deep Learning

Adds antialiasing (smooth edges) to an image containing black & white (1-bit) drawings and text.

Here are some examples, which are ideally viewed without browser scaling. They show the B&W image, followed by the algorithm output, followed by the ground truth (original image, with proper antialiasing):

Details

The algorithm uses a UNet based on ResNet18.

Running

Requires Python 3.7, pytorch 1.3.0, fastai 1.0, Pillow 3.1.

The script src/generate_images.py will generate training images in data/train.

Then you can use src/fai_learner.py to train the model and process the test images from data/test.