The Taub Faculty of Computer Science Events and Talks

Patch matching and local treatment of images has proven to be a successful strategy for problems such as image denoising, inpainting, super-resolution, image editing and the list could go on and on. Naturally, the size of the patch is a crucial parameter of the patch matching algorithm. Practically, working with large patches often results in a deterioration of reconstruction performance. This is due the curse of dimensionality -- an increase in the patch-size requires an exponential increase of the database, in order to guarantee the existence of meaningful matches. Recently, a method for bypassing the curse of dimensionality was introduced by Romano et. al. The key in their method is the use of context feature that encodes the larger environment of the smaller patch, thus bypassing the disadvantage of working directly with bigger patches, while benefitting from the additional information they provide. The proposed method has been demonstrated on several image processing tasks that rely on patch-matching, showing the potential improvement in using their modified similarity measuring method induced by the context feature. In our research we improve upon these results by replacing the heuristic choice of the context features with learned ones. We investigate a deep-learning based strategy that preserve the order between distances and show that the order-preserving approach can improve upon previous work by a significant margin. On the theoretical front, we study the conditions for obtaining an ideal patch-ordering. Finally, we test the effectiveness of the proposed method on non-local means image denoising, depth image super-resolution and age estimation.

*Msc seminar under supervision of Prof. Michael Elad