On connectionism and the problem of correspondence

Anupam Joshi, Purdue University

Abstract

This dissertation deals with computational vision. It is argued that to come up with solutions for problems in computer vision, human vision must be set as an ideal, and that these solutions should use available information from neurobiological and psychological studies on human vision. The specific problems addressed in this dissertation are that of modeling the retinal processing and obtaining correspondence. Relevant information from human vision is presented and used to come up with the methods for their solution proposed in this thesis. First, an elegant and simple model of the processing done by the X type ganglion cells of the retina is presented. Assumptions which enable a modeling using layered feedforward networks are introduced and justified based on evidence from neurobiology. It is shown that the model learns to approximate Marr's operator, a commonly accepted definition of retinal processing. Experimental results are presented showing the operation of the model on grayscale images, as well as the speedups obtained when the model is implemented on various parallel platforms. Next, two new algorithms for correspondence are presented. One operates on conventional 2D images, the other on range images. An extensive case is made justifying the use of 3D data not just from a computer vision point of view, but also from the point of view of human vision. Both algorithms share a common approach to the problem, which involves computing approximations to motion parameters in order to obtain correspondence. This approach is justified from psychophysical and neurobiological perspectives. The algorithms, unlike much of the extant work, do not make too many simplifying assumptions, and are suited for work on real images. Extensive simulation results are presented, using both synthesised and real images, to establish the voracity of the algorithms. The reader is also introduced to related problems worthy of further exploration and insights, gained in the course of this work, that will be of help in solving them.

Degree

Ph.D.

Advisors

Lee, Purdue University.

Subject Area

Computer science|Ophthalmology

COinS