Face Recognition Based On Muilti-wavelet Features

With the rapid development of information industry, face recognition has been studiedwidely by researchers because it is more consistent with human cognitive habits thanfingerprint, iris recognition, etc, and it is directly and friendly in collecting and identifyingprocess. However, the changes of illumination, expression, pose and shielding are some veryimportant factors which affect face recognition. To reducing the affects of variableillumination, the images with different illumination conditions are analyzed. It can be seenthat the pixel difference between different points is stable in the same image. Motivated bythis, a face recognition method based on the multi-wavelet and Grayscale Arranging Pairs(GAP) is proposed, in which the arranged pairs with stable pixel difference in the lowfrequency multi-wavelet components are extracted as the face features. In addition, the sparserepresentation can solve the illumination, expression, occlusion problem very well because anundetermined face image can be sparsely represented with some images of training set byprojecting it to the training images. So, a face recognition method based on the sparserepresentation and multi-wavelet is proposed, in which the low frequency multi-waveletcomponents are extracted to form a redundant dictionary.The detailed work is as follows:1) Image enhancement method in logarithmic domain. According to the characteristics offace images with variable light, the images are preprocessed with the logarithmicnormalization method in this paper.2) Multi-wavelet feature extraction. By studying the multi-wavelet features of face imagein different frequency bands, it can be see that the multi-wavelet low frequency features canrepresent the main information of a face image well, and the high-frequency information oftencontains noise. Therefore the low-frequency information of a face image is extracted as thepreliminary features.3) An approach based on multi-wavelet and GAP is proposed. First, we obtain the stablepixel difference points in the multi-wavelet low frequency preliminary features. Then thebackground templates are built. At last, the classification is performed by comparing thematching degree between the stable pixel difference point pairs of the multi-wavelet lowfrequency sub-image of a test image and the background templates.4) A method based on multi-wavelet and sparse representation is proposed. We use thelow frequency sub-band information extracted by multi-wavelet to form a redundantdictionary, and the K-SVD dictionary learning algorithm is used to optimize the redundant dictionary. The test sample is recognized by calculating its sparse representation inredundancy dictionary.In this paper, the experiments are designed separately for two proposed methods. First,the experiments on the extended Yale B and CMU PIE face database are performed with themethod based on multi-wavelet and GAP, in which the illumination changes significantly.Second, we do the face recognition experiments on the Yale and AR face database using themethod based on multi-wavelet and sparse representation, which has variable illumination,posture, facial expressions and shielding. The experimental results show that both the twoproposed approaches can effectively recognize faces and have strong robustness.