Unconstrained handwritten digit VLSI recognition system based on combined neural networks
Guoxing Li, Bingxue Shi · 2002
A totally unconstrained digit recognition system based on cellular neural network (CNN) and multilayer perceptron (MLP) and its analog-digital mixed mode VLSI implementation is presented in this paper. The CNN is used to extract CCD (connected component detector) feature from 24/spl times/24 normalized digit image in horizontal, vertical directions and two diagonal lines row by row. These features are fed into the two layers MLP in time sharing mode after proper compression. MLP has 60/spl times/20/spl times/10 structure with local connection. It consists of 10/spl times/10 neural processing unit (NPU) array which contains latches, weight generating circuits and computation circuits, switched current integrators, current threshold unit and control logic block. The weights are programmable and their control codes come from off-chip EPROM. This recognition system is very smart and effective, it can be implemented in standard digit CMOS technology.