 |
Center for Nonlinear Studies |
Images from natural scenes have a lot of structure: they contain objects of all sizes, textures or uniform patches, edges and contours, etc. The statistics of natural images is not only important in computer vision, with applications to image compression, denoising or classification, but it also plays a crucial role in understanding the design principles underlying neural sensory coding. Natural images share some qualities with critical systems in physics, the most striking one being perhaps scale invariance. We explore further the connection between natural images and critical systems by proposing an explicit mapping between images and spin systems, and by building the thermodynamics of natural images from small patches of increasing size. We find behaviors reminiscent of a second-order phase transition, and the analysis of the (rugged) energy landscape suggests a novel neural coding scheme consistent with known response properties of neurons in the visual cortex.
Host: Lenka Zdeborova