面向图像识别的条件互信息特征选择方法  

Feature Selection Method Based on Conditional Mutual Information for Image Classification

作　　者：金聪[1] 金枢炜[2] 

机构地区：[1]华中师范大学计算机科学系,湖北武汉430079 [2]武汉大学物理科学与技术学院,湖北武汉430072

出　　处：《测试技术学报》2010年第5期459-462,共4页Journal of Test and Measurement Technology

基　　金：湖北省自然科学基金资助项目(2008CDB349);华中师范大学中央高校自主科研资助项目(CCNU09Y01013)

摘　　要：提出了一种基于条件互信息的图像特征选择方法.为了预测条件互信息,该方法选择与已选定特征具有最大熵的那些特征,并将选择出的特征进一步用于数字图像识别.图像识别器由支持向量机实现.实验中,识别器的输入数据是由人脸和非人脸组成的二类图像,这些图像均为大小是28×28且具有256个值的灰度图像.本文不仅将新方法用于图像识别,而且还将新方法与已有的识别方法,如经典的贝叶斯理论、神经网络、kNN等进行了比较.实验结果表明:新方法不仅能够在较短的时间内实现图像特征的选择,而且对图像识别有着比其它方法更高的正确识别率,完全可以用于图像识别.A novel image feature selection method based on conditional mutual information was proposed. For predicting conditional mutual information, the method selected those features which has the maximum entropy with the features already picked, to be used for digital image classification. We implemented an image classifier by SVM. In the experiment, the input data of the classifier were grayscale images with the size of 28× 28 and the 256 grayscale levels, and covered face and non-face images. We compared several image classification methods including the proposed method, the Bayesian theory, neural network and the kNN. The experimental results show that the method requires less time and higher correct recognition rate than other classifiers.

关 键 词：特征选择 图像识别 互信息 SVM 识别率 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]