Study On Face Recognition Based On Partial Least Square Algorithm

Due to the high dimension of the face image, the method of processing directly inthe original image will increase the complexity of the algorithm, and it is also achallenge to the computer’s performance. The key to solve the problem is how to extracteffective identification features. This paper focus on the subspace statistic methods,which can be applied to face recognition, and some kind of good results are expected tobe obtained.Partial least square, a multivariate statistical algorithm, has been widely used inface recognition in recent years. Compared with other algorithms, it has deficienciessuch as lacking of nonnegative and sparse. In order to overcome the shortcomings of thepartial least squares algorithm, the paper does the research by two improving ways:adding nonnegative and sparse constrains. The main results can be summarized asfollows:1. Traditional subspace statistic methods, such as principal component analysis(PCA), only learn holistic, not parts-based, representations which ignore the availablelocal features (eyes, nose) for face recognition. However, these methods incorporatingthe category information such as linear discriminant analysis (LDA), faced smallsample problems. The paper proposes a novel approach to extract the facial featurescalled two-dimensional nonnegative partial least squares (2DNPLS).The main idea ofthe approach is grabbing the local features via adding the constraint of nonnegative to2DPLS, which makes the approach gaining not only the advantages of2DPLS,incorporating both inherent structure and category information of images, but also thelocal features, having nonnegative interpretability. For evaluating the approach’sperformance, a series of experiments are conducted on two famous face imagedatabases ORL, Yale face databases, which demonstrate that the proposed approachoutperforms the state-of-art algorithms.2. When there exist noises or occlusion the performance of face recognition basedon partial least squares is worse. Researches show that noise variables entering the PLSregression via direction vectors by iteration and imposing sparsity on the directionvectors can solve this problem. In order to improve the robustness of2DNPLS, thepaper proposes a novel approach called two-dimensional nonnegative sparse partialleast squares (2DNSPLS), which imposes sparsity on the direction vectors. Experiments show that the algorithm has good robustness for the image with occlusion, and therecognition rate is superior to other algorithms.