Analysis of image sequences using redundant representations

Citation

Rakshit, Subrata (1994) Analysis of image sequences using redundant representations. Dissertation (Ph.D.), California Institute of Technology. doi:10.7907/6vw2-a361. https://resolver.caltech.edu/CaltechETD:etd-10172005-130626

Abstract

This thesis focuses primarily on two techniques often used for low level analysis of motion in image sequences, the computation of optical flow fields and the determination of the spatio-temporal frequencies present. The complexity, cost and accuracy of image sequence processing is shown to be related to the manner in which the information is represented. The optical flow problem is formulated in terms of the basis functions underlying the discrete representation of images. Besides giving good results for a wide variety of inputs, this formalism highlights the benefits of a redundant representation of the input image in terms of reducing the overall cost of analysis. The spatio-temporal filter banks are analyzed using information theory. This approach makes explicit use of the input prior and provides an objective way of comparing the cost effectiveness of filter banks of different sizes. The output is used to generate a probability distribution over selected parameters in order to provide higher visual modules with a richer input. The formalism developed here provides a means of measuring the redundancy in filter bank outputs. This redundancy is shown to provide robustness to noise within the system. Likewise, a redundant representation of the image can also be used for error correction in case of corruption during storage or transmission. An algorithm for using the Burt Laplacian pyramid for such error correction is also developed and demonstrated.

Thesis Files