Sparse codes for natural images

Abstract

The human visual system, at the primary cortex, has receptive fields that are spatially localized, oriented and bandpass. It has been shown that a certain learning algorithm to produce sparse codes for natural images leads to basis functions with similar properties. This learning algorithm optimizes a cost function that trades off representation quality for sparseness, and searches for sets of natural images, which basis functions lead to good sparse approximations. The result of the learning algorithm is a dictionary of basis functions with localization in space, direction and scale. In this paper, dictionaries for different set of images are showed and their own properties are described and verified. It will be showed that the learning algorithm leads to overcomplete bases functions that “capture” the intrinsic structure of the images. This allows efficient coding of the images with good representation quality. The results are applied to image approximation and denoising.