Modelling early stages of human vision

摘要

This report presents the first results of an effort to model vision processes operating in the early layers of the human retina. The ultimate goal of this research is to develop a robot vision system based on computational principles of human vision. These include massive parallelism, dynamic feedback, and multilayer pattern recognition. Two neural models were developed. The first represented early vision in terms of static two-dimensional linear equations using a linear matrix and limited feedback. The second considered dynamic two-dimensional nonlinear processes and used matrices of nonlinear differential equations. The first set of equations was parameterized using psychophysical data from subjective intensity judgements for visual illusions. Illusions were quantified by creating a digitized image representing the subjective effect of human perception when exposed to input patterns. A Fourier transform was made of the input and output patterns and used to solve for the coefficients of the retinal operator equations. The equations were then applied to new illusions and compared with human subjective results through three dimensional plots of output pixel intensities. The second model with dynamic equations was studied using a computer simulation developed to operate on varying input wave forms and variable neural connection topologies. 27 refs., 11more » figs., 5 tabs.« less