ICA mixture models for image processing

We apply a probabilistic method for learning efficient image codes to the problem of unsupervised classification, segmentation and de-noising of images. The method is based on the Independent Component Analysis (ICA) mixture model proposed for unsupervised classification and automatic context switching in blind source separation [I]. In this paper, we demonstrate that this algorithm is effective in classifying complex image textures such as trees and rocks in natural scenes. The algorithm is useful for de-noising and filling in missing pixels in images with complex structures. The advantage of this model is that image codes can be learned with increasing numbers of basis function classes. Our results suggest that the ICA mixture model provides greater flexibility in modeling structure and in finding more image features than in either Gaussian mixture models or standard ICA algorithms. 1 Learning efficient codes for images The efficient encoding of visual sensory information is an important task for image processing systems as well as for the understanding of coding principles in the visual cortex. Barlow 121 proposed that the goal of sensory is to transform the input signals such that it reduces the redundancy between the inputs. Recently, several methods have been proposed to learn image codes that utilize a set of linear basis functions. Olshausen and Field 131 used a sparseness criterion and found codes that were similar to localized and oriented receptive fields. Similar results were obtained by Bell and Sejnowski [4] and Lewicki and Olshausen [5] using the infomax ICA algorithm and 8