Real-time early vision neurocomputing

Summarizes recent work on real-time implementations of early vision neural computations on the PIPE video-rate parallel computer. PIPE is an eight-stage machine for real-time image processing, performing nearly one billion 8-bit operations per second. Temporal computations are implemented using the integral solutions of the differential equations that govern the neurodynamics via an exponentially fading memory. Spatial computations are performed by convolving the images with masks. Dynamic neural networks that perform light adaptation, spatial contrast enhancement, and image feature velocity extraction have been implemented to process live imagery (256*256 pixels, 8-bit data) at 30 frames/sec. These computations simulate spatiotemporal processing in the retina and the motion pathway of the brain. Some examples of these computations are described.<
>