摘要
在手写数字图像的特征提取中,提出一种结合Fisher线性判别的多分辨率Gabor滤波方法,在所有特征点上寻求特定滤波方向上的局部最优滤波频率,以获得最佳滤波效果,同时压缩不相关特征。在MNIST手写数字图像库上的识别实验表明:在小样本情况下,该方法能更准确地抽取手写数字图像特征,识别效果明显优于直接进行Gabor特征提取。
In extracting the features of handwritten numeral, a method bases on multi-resolution Gabor filter which combined Fisher＇s linear discriminant is presented. According to the method, local optimum filtering frequencies in certain orientations is determined at all of the feature points. The purpose is to obtain the best filtering results and compress the irrelevant features. Experiments for handwritten numeral recognition to MNIST database indicates that the method can extract the features of handwritten numeral more efficiently and the effect of recognition is obviously better than directly extracting Gabor features conditions of small sample.
出处
《计算机工程与设计》
CSCD
北大核心
2007年第8期1870-1872,共3页
Computer Engineering and Design
基金
教育部博士点科研基金项目(20030532004)
