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.
- Implement a CLI.
- Find a better heuristic for patch approximation order.
- Rewrite the model in PyTorch for GPU acceleration.
- Make the linear model a parameter of the Inpainting class.
- Bin Shen and Wei Hu and Zhang, Yimin and Zhang, Yu-Jin, Image Inpainting via Sparse Representation Proceedings of the 2009 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP ’09)
- Julien Mairal Sparse coding and Dictionnary Learning for Image Analysis INRIA Visual Recognition and Machine Learning Summer School, 2010
- A. Criminisi, P. Perez, K. Toyama Region Filling and Object Removal by Exemplar-Based Image Inpainting IEEE Transaction on Image Processing (Vol 13-9), 2004