A systolic neural network image processing architecture

Abstract

The architecture introduced, the neural accelerator board (NAB) system, has been designed to perform a range a machine vision functions, of which one is motion parallax range estimation. The approach used in the NAB system, is to perform real-time calculations on the image data, captured by a moving video camera, using a highly parallel neural network architecture. The NAB system architecture, its physical implementation, and the surrounding application system are described. The implementation of the motion parallax algorithm, as a demonstration of the capabilities of this architecture, is outlined.<>