基于小波变换和神经网络集成的笔迹鉴别方法*

在笔迹鉴别中为了便于获取特征字符的细微特征,基于线性矩和小波变换提出了提取特征字符纹理特征的方法.小波变换能有效地提取字符的结构特征,而矩能够很好地对其进行描述.在该方法中,一幅特征字图像可以用一个含有52个元素的特征矢量表示,然后通过训练多个神经网络,并应用神经网络集成的方法将其结果合成,对特征空间进行正确分类.分别在特征字和候选人数变化的情况下进行实验,实验结果显示识别准确率较同类算法平均提高百分之五.

Handwriting verification using wavelet transform and neural network ensemble

In order to obtain the exiguous character of handwriting, texture feature of single character was extracted based on linearity moments and wavelet transform. The feature of character construction could be extracted by wavelet transform effectively and described by moments. A Chinese character image could be compressed into a vector of 52 elements in handwriting verification system. A finite number of neural networks were trained, and the method of neural network ensemble was adopted to combine their results for classifying the space of the feature vector. The experiments were done based on different numbers of feature characters and candidates, and comparative experiments were also done with the methods proposed by other researchers. Experimental results show that in average 5% improvement of recognition accuracy can be obtained, which indicates the method can be put into handwriting verification effectively.