摘要
本文提出了一种基于细胞神经网络(CellularNeuralNetworkorCNN)和多层感知机(MultiLayerPerceptronorMLP)构成的组合神经网络进行手写体数字识别的系统及VLSI的实现.该系统首先由CNN从归一化为20×20的字符图像的水平与垂直方向以及两个对角线上逐行进行相邻单元检测(ConnectedComponentDetectororCCD),并把检测到的特征进行适当压缩,然后按一定的时序馈送到下一级的MLP来进行识别.MLP是一个80×20×10的具有局部互连的两层网络.它主要由神经处理单元(NeuralProcessingUnitorNPU)阵列、开关电流积分器以及电流比较器等构成.NPU的权重是可调整的,其值由反向传播学习算法事先学习好,再经过八值量化后每一个权重由四位来表示.该识别系统较为简单有效。
A VLSI implementation oriented off line handwritten digit classification system based on cellular neural network and two layer perceptron is presented in this paper.The cellular neural network is used to extract the connected component feature of normalized 20×20 handwritten digit image.These features are fed into the two layer perceptron after proper compression in time sharing mode.The two layer perceptron has a structure of 80×20×10 with local connection and it mainly consists of a weight programmable NPU array,switched current integrator and current comparators.This classification system has a smart structure and it can be directly mapped into current mode circuits whose technology is compatible with standard digital CMOS VLSI.
出处
《电子学报》
EI
CAS
CSCD
北大核心
1999年第11期143-145,共3页
Acta Electronica Sinica
基金
国家自然科学基金资助项目!(No .6 96 3 6 0 3 0 )
关键词
细胞神经网络
多层感知机
手写体数字识别
VLSI
cellular neural network
multi layer perceptron
handwritten digit recognition
switched current
VLSI
