Center-symmetric Local Difference Pattern And Its Application In Face Recognition

Face recognition has become a hot research topic in the field of pattern recognition and artificial intelligence. Though pleasant results have been achieved during the last few decades, Face recognition technology is still faced with many serious problems in practical applications. Because the distribution of facial features are very similar, and the face itself is a flexible objects, changes of lighting, expression, posture, accessory and time brings considerable trouble for correct identification. How can effectively identify a large number of people and meet the requirements of real-time is an urgent issue for face recognition technology. Among them, feature extraction is the key issue of face recognition, because it directly affects the classification and decision-making process, that is, directly determines the recognition accuracy of face recognition system, therefore, focusing on robust and efficiency feature extraction methods has been an important process to improve face recognition technology.This thesis studied on Feature Extraction problem in face recognition, and the focus is local binary pattern(LBP). It is a very effective non-parametric texture descriptor that measure and extract local image texture features, and has received more and more attention and achieved great success in the field of face recognition because of its relatively simple principle, good illumination robustness etc. However, LBP features are so fine that resulting in noise-sensitive and the texture descriptors decreases recognition efficiency for high dimension. This paper analyzes two deficiencies described above, and the main work is as follows:(1) this paper proposes a face recognition method about the Center-symmetric Local Difference Pattern(CSLDP), which uses the center-symmetric nature and the principle of grayscale difference. It can be effectively extract contour features of face and facial organs contour feature by encoding the changes of four center-symmetric directions. In addition, Histogram vector dimension obtained by CSLDP is only one sixteenth that of LBP, thus improving the processing speeds significantly;(2) this dissertation proposes a face recognition algorithm based on Gabor wavelet and CSLDP was brought forward. First, Gabor wavelet kernel function respectively convolute with every one face image, and then apply CSLDP to analyze the above results, finally, the histogram of Gabor + CSLDP mode values represent the input face image. Extensive experimental results on some publicly available face databases show that CSLDP and have strong robustness in face recognition under uncontrolled conditions, CSLDP maintains high recognition rate as well as high processing speed, meeting the requirements of real-time applications.However, the proposed method has some limitations. Firstly, the proposed CSLDP method still exist an issue that the simple central pixel is selected as the threshold value. Further, it is not accurate that CSLDP uses a certain value to indicate change of direction. In the future, we may consider the mathematical probability theory to establish its associated membership function model for the change. Secondly, Gabor features belong to a high-dimensional feature. Although the combination of CSLDP reduced its dimension, the reduce amount is limited. Thus the follow-up study can be combined some effective dimension reduction measures. Finally, this paper just makes a preliminary exploration of feature extraction, but the classifier design is also the important part of face recognition system and optimizing classifier the future of a search direction.