多层分组神经网络的手写体数字识别

手写体数字识别的应用研究是字符识别中具有挑战性的课题．提出一种基于二进小波变换与多层分组神经网络的自由手写体数字的多分辨率识别算法．该算法包含二进小波变换的多分辨率特征抽取单元及多层分组神经网络分类器，与传统的完全连接的神经网络相比，该网络结构简单、输入节点少，并且由于网络分为子网结构，不同子网学习的是不同的特征映射值，某一子网不收敛不会影响到其他子网的收敛，网络鲁棒性好．采用信函分拣机提供的字库测试表明，其正确率为９８％左右．

Multiresolution Recognition of Unconstrained Handwritten Numerals with Multilayer Cluster Neural Network

The research on recognition of handwritten numerals is the most difficult and challengable problem in OCR. In this paper, a new scheme of multiresolution recognition of unconstrained handwritten numerals based on dyadic wavelet transform and multilayer cluster neural network is presented. The scheme includes a feature extractor with dyadic wavelet transform and a classification multilayer cluster neural network. Compared with the conventional neural networks, this kind of neural network is simpler and has less input nodes. As the network is divided into subnets and every subnet is trained according to different features, if one subnet does not converge, it will not influence other subnets. Consequently, the network is more robust than the conventional network. In order to test the validity, experiments with unconstrained handwritten numerals database provided by a mail sorting machine are made. The correct rate is about 98%.