Abstract

Neural vision models, both biological and computational, are characterized by highly parallel simple processing units, partially local and partially global connectivity, relatively slow response times, and highly parallel input/output. We explore potential uses of optics in some neural vision models, with emphasis on early vision. Early preprocessing and feature extraction stages typically utilize interconnections in a moderate sized local neighborhood, that are fixed and largely space-invariant. A useful and biologically motivated example is the extraction of Gabor transform coefficients from a set of receptive fields. A sample application of optics to the implementation of retinal ganglion cell and simple cell operations of the visual cortex are discussed. This includes an experiment utilizing fixed space invariant holographic interconnections, and an incoherent optical neuron (ION) model implementation of bipolar neuron units using liquid crystal light valves. More complex neuron functions, such as Grossberg’s mass action law neuron units useful for input normalization, can also in principle be implemented utilizing either the ION model or optoelectronic array devices. The next higher set of vision levels incorporates a larger degree of space variance, and performs such operations as higher level feature detection and partial invariance to distortions of the input such as translation and rotation. More computationally oriented neural vision models can provide subsequent levels, and typically utilize more global connectivity and adaptive interconnections. Examples include von der Malsburg’s dynamic link architecture and Chellappa’s unified vision models. Optics may play a critical role in the implementation of some of these models.

© 1991 Optical Society of America