NeNEB-an application adjustable single chip neural network processor for mobile real time image processing

Abstract

A digital implementation of the recall phase of a backpropagation neural network for real time image classification is presented. The motivation has been, that parallelism of a neural network has less advantage, if the input data stream is sequential such as the pixel stream of an usual CCD camera. In addition, classifying a stream of pixels with a single chip neural processor implanted into a camera avoids the bottleneck caused by image data transfer and storage, which is neglected often. But the chip will not be restricted to real time image processing applications. The chip realizes a network with 32 output neurons, 8 hidden neurons and up to 64 K inputs (/spl ap/512 K synapses). It is estimated to classify 50 gray scale images per second of size up to 256/spl times/256 pixels, which complies to 26 M cps. The weights of hidden neurons are stored in external memory, whereas weights of output neurons are stored on-chip.