An Illumination Independence Face Recognition Algorithm Based On Gabor And SVM

Because of the broad application prospects of face recognition in the public security department, security verification system, human-computer interaction, credit card verification,it will become the study hotspot in the field of pattern recognition and artificial intelligence. As the influence in light, gesture, facial expressions and age, face recognition system performance dramatically.This paper focuses on the perspective of light,studys how to improve recognition rate on the face image,which is uneven illumination, light affecting seriously.This paper presents a wavelet-based face image preprocessing algorithm for illumination.Firstly,2D discrete wavelet transform(DWT) decomposes a face image into four subband frequency,which is the low-dimensional approximation(LL) coefficient of the details of components and high-frequency detail coefficient(LH,HL,HH). Secondly, it convert the LL component into logarithmic transformation,the LH, HL and HH components are attenuated respectively. Finally, perform inverse of DWT on all the coefficients in wavelet subbands. Experiments show that the algorithm is simple, fast speed, image processing illumination effect.Gabor wavelet transform is not sensitive to light, in order to reduce the computation of the convolution of GWT, this paper proposes a face feature extraction algorithm based on meaning GWT and meaning grid block.Firstly,it generates 40 Gabor filters(5 scales and 8 directions) to meaning Gabor filters.Secondly,filting the face image with the eight meaning Gabor filters and mean sampling characteristics have been filtered image with 4×4 grid blocks.After together all the characteristics, get the whole picture of Gabor feature vectors.Finally, the dimensionality reduction process with PCA/LDA method for face feature vector.Proposed a classification algorithm(NN-SVM algorithm),which combining nearest neighbor(NN) and support vector machine(SVM). Firstly, it calculates the distance and the number of votes between each testing sample and training sample with NN and SVM classification algorithm respectively.Secondly, for each test sample, were selected from the smallest dimension vector K1 and K2-dimensional vector of the greatest number of votes, the joint formation of a new K1+K2-dimensional vector. Finally, as a unit row vector,calculates the vector value of the maximum frequency, that is, the sample NN-SVM algorithm classification results. In this paper,Yale B face database as experimental subjects,subset 1 is used for training and the subset 2,subset 3,subset 4 and subset 5 are used for testing.Results shows that the proposed method can achieve 100%, 97%, 95% and 95% recognition rate in subset 2 to subset 5 respectively.