Single Sample Face Recognition Research Based On Local Direction Pattern

Face recognition is a pattern recognition method, a non contact, portability, difficult to change and other unique advantages, in the field of identity identification and verification by more and more attention and focused by image processing and machine vision field. After many years of hard work, the recognition rate of face recognition algorithms at home and abroad is getting more and more efficiently, which can reach the practical stage. However, the current national citizen identity information has the only a photo on passport, in the public security case, customs inspection, monitoring and investigation needs to determine the identity of the face recognition applications, most algorithms have serious challenged. So, only a sample of the face recognition, becomes a research focus in the field of face recognition.The texture is widespread in nature, is the common characteristics of surface,a visual feature does not depend on the change of color or brightness, describing the image pixel gray level in the neighborhood spatial. Texture analysis is an important research topic in computer vision and image processing, and how to get the texture feature which is one of the important links.The method of local texture feature extraction based on Local Binary Pattern(LBP) plays a significant role In the face recognition, it has the advantages of simple calculation, the effect of light is weak etc. But in the complex background, noise, occlusion or extreme complex conditions, the classification efficiency of LBP is greatly reduced. To overcome the limitations of LBP,a improved algorithm was proposed. The main research work and innovation contributions are follows:(1) The face images are convolution with the primary derivative of the Gauss function to overcome the influence of illumination and noise. The gradient components of the image were obtained, and the gradient magnitude and gradient phase were calculated. The influence of external environmental factors on human face images is reduced.(2) Two improved feature extraction methods are introduced, the ?-LBP(? Local Binary Pattern) operator and the center-symmetric local binary pattern(CSLBP) which increased the applicability of the traditional LBP operator. The ?-LBP operator can extract features in different subspaces, which increased the sample diversity; CSLBP inherited all the advantages of LBP, but the amount of computation is reduced by half.(3) The previous scholars only consider the amplitude, but ignore the impact of the phase on the classification. The gradient phase was quantified, and then the gradient phase and gradient magnitude were performed by using the CSLBP. Then the three codes were concatenated into the general feature vector for face recognition. The classification ability is greatly improved.(4)Finally, the feature maps are divided into several blocks and the concatenated histogram calculated over all blocks is utilized as the feature descriptor of face recognition. The recognition is performed by using the histogram intersection classifier. Experimental results on AR and CMU-PIE face databases validate that the new algorithm is an outstanding method for single sample face recognition under different illumination conditions, different facial expression conditions and partial occlusion conditions.