Research And Realization Of Face Recognition Algorithm Based On DCV And SVM

After several decades of development, face recognition have become a hot research topic in the pattern recognition field and received extensive attention more and more. Although face recognition technology has developed rapidly in recent years, it still not well applied in actual commercial products。The "Small Sample Size" problem is one of the obstructions of the face recognition. To solve this problem, this paper gave a further research on the DCV (Discriminative Common Vector) Algorithm, and proposed a way for face recognition based on the feature extraction algorithm of DCV and SVM (Support Vector Machine, SVM) classifier.DCV is a new feature extraction algorithm which developed only in recent years, it is based on FLDA (Fisher Linear Discriminant Analysis). DCV algorithm can retain all the optimal discriminant vectors of FLDA. Besides, it can overcome the pathological singular solution when FLDA is used to resolve the "Small Sample Size" problem. The thesis use the typical nearest neighbor algorithm as the classifier, and give experiments on the PCA, DCV and the PCA+DCV feature extractions. By comparing the performance of this three algorithm, The thesis got a set of conclusions:In small samples of the feature extraction, DCV algorithm can get higher recognition rate; the2nd DCV algorithm has the best stability; Although the2nd DCV algorithm and PCA+DCV algorithm have the same effect on feature extraction,the PCA+DCV algorithm is more time-saving than2nd DCV algorithm. Since nearest neighbor classifier performance have much limitations, the DCV+SVM algorithm is proposed in this paper, here the thesis use the2nd DCV algorithm. The results show that when the number of training samples is2,3,4and5, the Recognition rate of DCV+SVM algorithm is7.81,4.64,5.00and5.00percentage higher than PCA+nearest neighbor classification algorithm respectively, and6.56,3.57,2.5,4.5percentage higher than PCA+nearest neighbor classification algorithm. This proves that the issue of small sample SVM has better classification ability when face to the "Small Sample Size" problem. At last, we load the DCV+SVM method on DSP image processing system. After testing this algorithm, we can get the conclusion that the recognition rate is stable whatever on the Matlab or the DSP, and the DCV+SVM method can meet the real-time requirement.