手写数字识别中的Gabor参数优化设计  

Optimum design for Gabor Parameters in Handwritten Numeral Recognition

作　　者：刘琼[1] 周慧灿[1] 王耀南[2] 

机构地区：[1]湖南文理学院计算机科学系,湖南常德415000 [2]湖南大学电气与信息工程学院,湖南长沙410082

出　　处：《微计算机信息》2007年第19期290-291,306,共3页Control & Automation

基　　金：教育部博士点科研基金(20030532004)

摘　　要：提出一种多分辨率Gabor滤波器组的参数优化设计方法,该方法通过分析在小波框架下的频率和带宽的相邻关系,推导滤波器组的参数构成,以尽可能少的滤波器覆盖尽可能大的信号频率空间;并采用两级、多通道2DGabor滤波器组进行数字图像特征抽取,然后结合多类SVM分类器进行分类识别。对MNIST手写数字图像的识别实验表明:在小样本情况下,该方法具有很强的特征抽取能力和较高的识别率。A new method of optimum design for multi- resolution Gabor filters in handwritten numeral recognition is put forward. By analyzing the adjacent relations of frequencies and bandwidths in wavelet frame, the composing of parameters in Gabor filter bank has been made a deduction. The purpose is to cover the frequency spatial as far as possible with the least number of Gabor filters. To extract features of handwritten numeral image, a set of 2D Gabor filters with two levels and multi channels is constructed. And then the features is classified and recognized by dint of a multi- class SVM classifier. Experiments in MNIST database indicate that the new method has strong ability to extract features and extremely high recognition rate conditioning small sample.

关 键 词：手写数字识别 2D GABOR滤波器 多类SVM分类器 特征抽取 

分 类 号：TP391.43[自动化与计算机技术—计算机应用技术]