Face Recognition Based On KPCA And LBP

Face recognition have already become a hot spot and tough task of research in digital image processing and pattern recognition, its application has shown an increasing importance in society. We can divide face recognition into three domains according to data types, which are face recognition based on static image, digital video and sensors like thermal infrared sensor and three dimensional data [5]. For the reason that human face is the most nature and direct biological feature than other feature such as fingerprint, iron and palm feature, it is unreplaceable. Although after nearly half a century of development, the performance of face recognition system is far from mature both in recognition rate and stability simply because non-rigid human face is susceptible to illumination, pose, occlusion and other undesirability circumstance in real application. Exploring a face recognition system that can conquer all these unsatisfactory and complicated facts is a subject with challenge and huge application prospect.LBP(Local Binary Pattern) is widely used in texture analysis, face recognition, face detection and facial expression recognition as its simplicity in calculation and strong power in description. However, cascading histogram dimension of LBP histogram is too large and sparse when block number is large, existing dimension reduction methods of LBP histogram, for example, rotation invariant LBP operator, bring down the description power of LBP operator and dimensions after reducing is not flexible. On account of these problems and traditional PCA have the deficiency of dealing nonlinear problem, utilization rate of limited samples is not high, this paper fused LBP histograms using different coefficients of limited sample images, and then reducing dimensions using KPCA method and researched LBP patterns under nonlinear subspace. The paper research content is as follows:1. Systematically introduces the face recognition, including an overview of face recognition, background, significance, development status and direction, test databases and classic face recognition algorithms, etc.2. On account of utilization ratio of samples is not high when traditional LBP diverse samples as sample library, this paper proposes a method called block LBP histogram weighted fusion method. The fused histogram and the traditional histogram of samples are putted together to constitute sample set. This paper also analyzes the impact of weighting coefficients to recognition rate.3. This paper studies the method of principle component analysis and the method of kernel principle component analysis and gives recognition performance on the general face databases. Furthermore, this paper analyzes the reasons of weak power of traditional PCA in handling nonlinear problems and uses kernel principle component analysis to reduce the dimensions of block LBP histogram. As extraction of energy value will lead to the difference of characteristic numbers when the dimensions of the block LBP histogram is reduced, this paper adopts the method of fixed numbers of characteristic value to analyzes the impact of different kernel functions, its parameters and the reduced dimensions to recognition rate and obtains the corresponding conclusion. The conclusion shows that this method is obviously better than PCA, LDA, KPCA face recognition methods in the matter of recognition rate.4. Researched recognition results under various similarity distances, such as euclidean distance, spearman distance, cosine distance and so forth, and obtains the best distance matching measurement for the feature after reducing dimensions.