Research And Implementation Of Face Recognition System Based On LPP Algorithm

In recent years, face recognition, widely used in fields such as Human-Computer Interaction, Information Security and Public Safety, has become the hot spot of research in the fields of pattern recognition and image processing.Face recognition includes face detection, face positioning, feature extraction, feature training, feature recognition and many other steps, the feature extraction refers to the extraction of low dimensional date which reflects nature characteristics of face from high dimensional data of face image, to simply the operation of data processing. The quality of the feature extraction method determines the face recognition method.LPP, which has the analysis ability of manifold, is a linear subspace learning method. It will achieve good results if LPP algorithm is applied to face recognition. However, LPP algorithm is an unsupervised learning method, and the recognition rate will drop significantly while the illumination, expression and posture of face images have great changes.This paper researches and realizes LPP algorithm primarily, then SLPP algorithm is proposed, aimed at the question that LPP algorithm makes incomplete use of class information from different face images in the process to obtain face subspace. Furthermore, the face recognition system based on LPP algorithm is finished:first read the face images, second preprocess face images by histogram equalization, scale normalization and difference of Gauss filter, then extract feature for face image with LPP and SLPP algorithm, lastly recognizes face images with nearest neighbor classifier. Finally, this paper talks about to realize parallelizing multi-thread OpenMP, parallelizing SSE instruction, multi-thread OpenMP combined with parallelizing SSE instruction in the system.This paper implements the face recognition system. The average recognition rate of SLPP tested on ORL face database reaches 88.49%, which increases 2.11% compared to the system using LPP. The average recognition rate of SLPP tested on Yale face database reaches 88.74%, which increases 1.6% compared to the system using LPP. The average recognition rate of SLPP tested on AR face database reaches 90.06%, which increases 2.38% compared to the system using LPP. This paper parallels the parts of feature extraction and feature recognition. The speed of multi-thread OpenMP parallelized feature extraction and feature recognition is 1.432 and 1.59 times faster than serialized system respectively. The speed of SSE instruction parallelized feature extraction and feature recognition is 1.276 and 1.362 times faster than serialized system respectively. The speed of multi-thread OpenMP and SSE instruction parallelized feature extraction and feature recognition is 1.745 and 1.932 times faster than serialized system respectively.