Research On Face Recognition Under Illumination Variations

Face recognition technology is an important research topic in finance, commerce, tourism and other fields, and has achieved significant research progress. But there are many unstable factors in the actual environment, such as illumination variation, pose variation, expression variation, making the face recognition technology facing the challenge. The illumination variation is a key factor which affect the face recognition technology. About the illumination variation, the main research contents of this paper are as follows:(1)The illumination normalization method is studied. By analyzing the illumination preprocessing chain method, the gamma correction is not good at dealing with face images under complex illumination, affecting the treatment result of subsequent Difference of Gaussian filter. To solve this problem, the log-domain Discrete Cosine Transform( Discrete Cosine Transform,DCT)is used as the first step of the illumination normalization method, and combined with the Gauss difference filter and the contrast equalization method. A log-domain illumination normalization method is proposed, and the experimental results show that the proposed algorithm can effectively reduce the effect of illumination on face image.(2)The local ternary patterns(local ternary pattern, LTP) algorithm is studied. The threshold of LTP is an empirical value, and plentiful experiments are needed to be made to find an optimal threshold. To solve this problem, the adaptive threshold local ternary patterns(adaptive threshold local ternary pattern, ATLTP) algorithm is proposed. First, the algorithm use the characteristic of standard deviation that the standard deviation can reflect the dispersion degree of sample variables, and the threshold of LTP value is estimated adaptively. Then, the texture features of face images are extracted. The experimental results show that ATLTP algorithm has stronger illumination robustness than the LTP algorithm.(3)The the multi-scale ATLTP algorithm is studied. The ATLTP texture feature is not rich and comprehensive, and the the multi-scale ATLTP feature extraction algorithm is proposed. In this method, first, the multi-scale ATLTP algorithm is used to extract different-scales texture feature from original face images. Then, the different-scales texture feature is divided into several units, and the histogram of each unit is computed. At last, different scales features are fused as facial features. The experimental results show that the multi-scale ATLTP feature extraction algorithm can effectively improve the recognition rate compared with the ATLTP algorithm.(4)The face recognition system of illumination variation is studied. there is not a kind of illumination normalization method which can eliminate the effect of illumination on face images, and there is not a kind of light illumination invariant feature extraction algorithm which can extract complete illumination invariant. To solve this problem, an algorithm for face recognition based on illumination normalization and ATLTP features is proposed. Before extracting ATLTP texture features, illumination normalization is used to reduce illumination effect on face images. The algorithm combines the advantages of illumination normalization and ATLTP algorithm, and experimental results show that the algorithm achieved satisfactory results.