Intel Collaborative Research Institute for Computational Intelligence
Understanding and Utilizing Natural Image Statistics
One of fundamental problems in computer vision and image processing is the learning of accurate image priors. Most tasks in image processing, computer vision and computational photography involve fundamentally ill-posed inverse problems, solved with the aid of image priors. Despite years of research, current image priors are only approximate, and mostly limited to local patches and kernels. A team of researchers from three universities (Technion, HUJI and Weizmann) attempts to make a fundamental contribution to the learning of more accurate global image priors and to their efficient application to a variety of image restoration problems. More specifically, this team intends to explore the interplay between local patch-based models, and global ones, the role of scale-invariance in such models, and fundamental limits to the performance of various inverse problems using patch-models.