A rotation-invariant embedded pattern recognition system

Abstract

Hardware implementations of neural networks (NN) offer superior performance over software implementations due to the inherent parallelism that can be exploited at the architectural level. However, they are rendered unsuitable for incorporation into low-cost pattern recognition systems, as they are expensive to implement in VLSI. This paper proposes a rotation-invariant embedded pattern recognition system based on an area-efficient NN architecture. It employs a novel feedforward multi-layered, time-multiplexed neural network architecture with multi-level threshold functions. Novel training and recall methods that best exploit this architecture have also been devised. Our results show that the proposed method exhibits superior learning and recognition abilities, and also tends itself well to a low-cost rotation-invariant pattern recognition system. Such a system can be used effectively for different industrial applications that involve machine vision or autonomous mobile robots.