Image inpainting via dictionary learning and sparse representation.
This project aims at rebuild "damaged" pictures by learning a sparse representation of non-damaged patch of the image.
The model is composed of 3 Linear regressions (one per channel) with L1 regularization (aka Lasso). It encodes the picture to a HSV color model, normalize its pixels between [-1, 1], and learn which sparse combination of pixels can properly rebuild the picture.