ANC/DTC Seminar: Yair Weiss, Hebrew University of Jerusalem (Host: Chris Williams)

Learning a statistical model of natural images is a longstanding topic of research in disciplines ranging from engineering to computational neuroscience. Somewhat embarrassingly, however, algorithms based on these statistical models do not give state-of-the-art performance in image restoration, and are outperformed by "block matching" methods that do not have an explicit probabilistic model. In this work, we show that many widely used statistical models of images are actually not very good density models and a simple, unconstrained, Gaussian Mixture Model (GMM) can give much higher likelihood to unseen images. Using the learned GMM, we obtain state-of-the-art image restoration performance and by examining what the GMM has learned we obtain new insights into the statistics of natural images that are not captured by most existing models.
Joint work with Daniel Zoran.