Neural networks are valuable tools for the support of a wide range of image processing applications. For video-rate operation, special-purpose parallel hardware is often necessary. One of the most common architectures used for this purpose is the linear systolic array. The design and implementation of multi-layer neural networks in linear systolic arrays can be complex, however. This paper demonstrates that the smallest network is not necessarily the best in terms of learning or recall times. Furthermore, this paper shows that the manner in which networks are mapped into a particular hardware structure affects both the performance of the application and the efficiency with which the hardware resources are used. We analyze and identify how to best structure neural networks to optimize network performance for throughput, latency and the efficiency with which the hardware is used. We use the HANNIBAL neural network processor as a research vehicle for these investigations and demonstrate the value of the proposed techniques by a number of example applications.

An analytical model is presented for assessing the performance of multilayer neural networks implemented in linear arrays. Metrics to assess latency, throughput rate, and computational and input-output bandwidth are developed. These metrics demonstrate a rich and complex interaction between the performance of the hardware and the number and relative dimensions of the layers in a network. Practical illustration of the use of these metrics is demonstrated for a two-hidden-layer network.>

The application of artificial neural networks to real-time image processing tasks requires the use of dedicated, high performance hardware. A linear array processor called HANNIBAL has been developed which implements the backpropagation neural learning algorithm on-chip. This paper considers the design of a complete neural system which integrates HANNIBAL into an existing image processing environment. The goals for the design of the system have been set partly by the primary application, namely feature recognition, but mainly by the desire for a flexible, high performance hardware tool for the study and evaluation of range of neural image processing applications.