Skip to main content
U.S. Department of Energy
Office of Scientific and Technical Information

Neural network modeling of early retinal vision processes

Conference ·
OSTI ID:6916838
This paper considers two techniques to obtain values for parameters in models of vision processes in early layers of the human retina. The work is part of a long-term effort to develop a robot vision system using three computational principles of neural networks. These principles are: massive parallelism, dynamic feedback, and multi-layer pattern recognition. Two network models were studied. The first represented lateral inhibition processes as a matrix of two-dimensional linear equations with limited feedback. The second represented multi-layer networks as matrices of dynamic, two-dimensional nonlinear differential equations. The lateral inhibition model was parameterized using psychophysical data from human judgments of pattern brightness in a machband illusion. A 2-D Fourier transform was made of original and perceived patterns and used to solve for the coefficient matrix of the neural network. The coefficients were then applied to the pixel matrix of a Herman grid illusion and compared with the corresponding human subjective results through three-dimensional plots of pixel intensities. The second model was analyzed using a new computer simulation developed to work with moving images and adjustable neural connection topologies. 19 refs., 8 figs.
Research Organization:
Oak Ridge National Lab., TN (USA). Engineering Physics and Mathematics Div.
DOE Contract Number:
AC05-84OR21400
OSTI ID:
6916838
Report Number(s):
CONF-8607100-1; ON: DE87000746
Country of Publication:
United States
Language:
English