Yair Weiss
Hebrew University
Learning to Perceive from Image Statistics- a Computational Challenge
Natural images take up only a tiny fraction of the space of all
possible NxN matrices. It thus makes sense that perceptual systems,
both biological and artificial, would use image statistics to improve
their performance. This idea goes back at least to Mach (1886) and
Helmholtz (1925) operationalizing it for computer vision has proven to
be difficult.
In this talk I will describe the computational challenges raised by
such an approach - learning very non Gaussian distributions in high
dimensional spaces and performing inference with such distributions. I
will then summarize some of our research in this direction. The good news
is that very simple statistical models can lead to surprisingly
powerful algorithms. The bad news is that even these simple
statistical models lead to complicated optimization
problems. Specifically, I will
discuss applications of image statistics to image denoising,
"inpainting" and transparency.
Joint work with A. Levin, A. Zomet and E. Levi.
Yair Weiss is a senior lecturer at the Hebrew University School of
Computer Science and Engineering. He received his Ph.D. from the
Massachusetts Institute of Technology and was a visiting scientist at
U.C. Berkeley. His research interests include human and machine
vision, machine learning and error correcting codes.