Continuous-time global computer vision with analog, specialized, and interacting neural networks

Abstract

Traditional computer vision considers early vision as an “inverse optics” problem and tries to invert projective and radiometric equations. It postulates independent modules and uses constraint satisfaction techniques within each module to obtain the desired inverse. We outline the shortcomings of these approaches and discuss how neural networks can overcome them. We review relevant findings from neurophysiology and psychophysics and indicate how they have been incorporated into neural network models. In particular, we stress massive parallelism, nonalgorithmic analog behavior, attention, goal-directed behavior, habituation, sensitization, self-organization, and local and global processing properties of neural networks as key elements to analyze visual inputs in nonstationary environments.