Discriminant uncorrelated locality preserving projection

Abstract

The basis vectors of traditional locality preserving projection (LPP) are statistically correlated. This makes the features extracted are redundant. In addition, LPP is an unsupervised feature extraction method because class information is not used in LPP. In this paper, a discriminant uncorrelated locality preserving projection (DULPP) algorithm is proposed. The DULPP overcomes the shortcomings of traditional LPP. It uses class information of training data when constructing the weighted neighborhood graph. The relationship among data can be described more accurately. Moreover, DULPP can extract features which are statistically uncorrelated. This can make the features extracted not only preserve the local information of original data space but also contain minimum redundancy. The experiment suggests that the proposed algorithm achieves much higher recognition accuracies. The proposed method can be used in video supervision system, target tracking and recognition system to pursue higher recognition accuracies.