Research Of Face Recognition Based On SIFT Algorithm

Face recognition has been a hot research field of biometrics, computer,(applied)mathematics, electronics, automation, visualization, virtual reality, image processingand pattern recognition, and other disciplines have a more extensive research, as wellas aerospace, weather, Criminal Investigation, borders, airports, check the field hasimportant applications. With the continuous development of the face recognitiontechnology, and how to improve the recognition rate of the face image under differentchanges, it has become a problem to be solved.Single image based face recognition under different variations such as light,expression, pose and occlusion, has been recognized as an important task in the realworld. The popularly widely used holistic features are easily distorted due to the light,or some other variations. In order to tackle this problem, recognition methods basedon feature descriptor have become more and more important, and have been obtainingperformance. The recently developed Scale Invariant Feature Transform(SIFT), whichdetects feature points sparsely and extracts feature locally for object matchingbetween different views and scales, can also benefit single image based facerecognition. However, because face recognition differs from generic object matching,thus SIFT should not be directly used for face recognition, there are still issues to beresolved:(1) most of the face recognition method treat face image as a generic objectimage, two face image matching, directly use the key points which detected by SIFTfeatures for matching between two face images, More precisely, these features are notdirectly for face recognition.(2) The computation of feature matching strategy between two images is toointensive, or the matching rate is not high.To this end, this paper developed a new framework for detecting feature keypoints sparsely, describing feature context and matching feature points between twoface images. We call this new proposed framework as Face Sparse Descriptor(FSD).This paper has the following main aspects:1. Study the principles and steps of the SIFT algorithm, by analyzing the SIFT aswell as face characteristics, this paper proposed a method named FSD which candetect feature points, then demonstrate the processing results of SIFT and FSD based on the single face image of AR, CMU, and FERET face database.2. This paper improved the feature descriptor generation steps of the basic SIFTalgorithm, Matching single face image under experiments, light, pose, and occlusion.For single image which has been normalized, FSD algorithm has higher recognitionand strong robustness compared with SIFT, Local Binary Patterns (LBP), Histogramsof Oriented Gradients (HoG) and Principal Component Analysis (PCA) algorithm.3.This paper improved the matching strategy based on the basic SIFT algorithm,after the generation of the feature descriptors, storage the detected feature points byR-table, calculate the average local distance of the descriptor and weighted the resultswhile matching, were instead of the SIFT matching the Euclidean distance weighting,used the weighted average local distance instead of the Euclid Distance, therebyreducing the probability of a mismatch points and receive more convincing overallmatching degree.Experiments are conducted on AR,CMU and FERET databases, comparing theresults of FSD, SIFT, PCA, Gabor, and other local feature descriptor(i.e, LBP andHoG) with the influence of a variety of conditions, the experiment results illustratethat the FSD method proposed in this paper is effective.