一种推广的K-L展开方法及其在人脸识别中的应用

A GENERALIZED K-L EXPANSION METHOD AND ITS APPLICATION IN FACE RECOGNITION

本文推广了从类平均向量中提取判别信息的K-L展开方法,使其适用于小样本问题的特征抽取,并从理论上较深入地探讨了高维、奇异情况下如何降低计算量的问题,给出了一个简单而高效的算法.在ORL标准人脸库上进行测试.由本文算法抽取的持征在最小距离分类器和最近邻分类器下均达到96%的正确识别率,这一结果大大优于经典的特征脸方法和Fisherfaces方法在该库上的识别结果.

The K-L expansion technique that is capable of extracting discriminatory information contained in class-mean vectors is a very effective method for linear feature extraction. However, it is not applicable in problems with small sample size, such as face recognition, because the within-class scatter matrix is always singular. In this paper, this technique is generalized to suit for problems with small sample size. And, to deal with the high-dimensional difficulty in face recognition problems, investigation is conducted on how to reduce the computational complexity in theory. As a result, a simple and efficient algorithm of K-L expansion for high-dimensional and singular case is developed. Finally, our algorithm is tested on ORL face image database, and a recognition rate of 96 % is achieved by using either a common minimum distance classifier or a nearest neighbor classifier. The experimental results also demonstrate that our method is superior to the classical Eigenfaces and Fisherf aces.