Learning to Deform for Efficient Image Compression

Image Compression algorithms aim to encode an image in a way that can reconstructed with minimal error. Common algorithms base on error measurements that are very sensitive to geometric shapes in the image. As a result, the compressor wastes many bits to encode perfectly geometric shape while the human eye will not always notice the difference.
Based on this, Ms. Tamar Roth Shaham proposed to deformation the picture so that on the one hand it will not be notice by the human eye and on the other hand will improve the compression capabilities of the various algorithms by matching the geometric shapes to the required. While the deformation is almost unnoticeable to the human eye, it has a great effect on the compressor and improves image quality after compression. In order to find deformation, the compression algorithm used as a "black box," so the innovative approach can work with any existing compression algorithm.

By being an iterative algorithm, its running time is long. Therefore, we would like to train a learning system that will study the algorithm and will be able to offer the relevant shift to each image according to the desired compression algorithm, in a very short time and without any harm to efficiency.
The project is a continuation of Tamar's work and part of Prof. Tomer Michaeli's research. Its purpose is to bring the results of the iterative algorithm presented in article [1] closer together by use of a convolution network.

Learning to Deform for Efficient Image Compression
Learning to Deform for Efficient Image Compression