Real-Time Image Processing Architectures for Perceptual Grouping, Depth Segregation, and Object Recognition.

摘要

The goal of this research program was to discover and develop real-time neural architectures capable of autonomously carrying out image processing and pattern recognition tasks in environments wherein noisy and unexpected events can occur. Such architectures are needed to cope with the fact that, in naturally occurring scenes, edges, texture, shading, size, stereo, and motion information are often overlaid and are viewed under variable illumination conditions. Special-purpose vision algorithms that can process only one of these types of information do not function well under naturally occurring conditions. The present work has analyzed a large body of data from visual psychophysics and neurobiology in order to discover and test neural principles and mechanisms whereby such a general-purpose competence is achieved by humans and animals. These designs are embodied in multi-level neural networks which are defined by novel types of nonlinear dynamical systems. The networks are computationally characterized for use both in explaining biological data about vision and pattern recognition, and in implementing novel image processing circuits for use in technological applications. Predictions of the theory have also been successfully tested in our psychophysics laboratory.