Ensemble Local Block Linear Discriminant Analysis For Face Recognition

In recent years, face recognition has been given more and more attention. Many research institutions and individuals have done lots of researches, and put forward some new recognition algorithms, but a large number of tests and practice show that a lot of problems need to be resolved in order to realize a really practical face recognition system. Nowadays, the performance of face recognition is mainly affected by expression, illumination, posture, age, background, image quality and computational complexity. This paper has got the following innovational achievements after studying the illumination and the computational complexity:(1) It proposes the method of Gabor features extraction and integration, based on uniform image block. Gabor wavelets have so good biological visual characteristics that it can extract a good face image texture feature, which is considered to be one of the best face representation methods. However, the dimension of Gabor features is too high to benefit the follow-up processing. This paper uses the way of uniform partitioning to group Gabor features, allows the overlap between blocks and uses principal component analysis and linear discriminant analysis classifier design for each group features, so it not only can reduce the feature dimension but also can improve the ability of the classification algorithm by ensemble. Considering the fact that different parts of the face have different classifying abilities, like the classifying ability of the important organs' image sub-block including eyes, nose, mouth, etc, is stronger than other sub-blocks, it is necessary to give different sub-classifiers different weights. The ensemble by using the ratio of the between-class differences to the within-class differences of each image sub-block feature vectors of training sample set matrix as the weight of each sub-classifiers in this paper;(2) It proposes the illumination preprocessing method of combining the logarithmic transformation and local normalization. Lots of experiments shows that light is the vital factor affecting the face recognition, and how to eliminate the the influence of illumination change on face image is also one of the hottest topic at present. Logarithmic transformation can expand the low gray area of image, compress its high gray area, and make the gray distribution of the whole image more evener so as to weaken the effect of illumination change; Local normalization is able to extract the features of facial image irrelevant to illumination, but it will be affected by the differences between high and low gray value in the partial windows in actual use. This paper analyzes the advantages and disadvantages of two kinds of preprocessing methods, puts forward the method of combing logarithmic transformation and local normalized preprocessing, and shows the validity of this method through the experimental results;(3) It proposes a new linear discriminant analysis algorithm which weightes the within-class convariance matrix. Linear discriminant analysis is a kind of facial feature extraction for the purpose of classfication, but the traditional linear discriminant analysis thinks that all face images make the same contribution to classification, without considering the influence of edge classes. To a certain degree, this paper enhances the algorithm's classifying ability in dealing with the edge classes through the way of weighting the within-calss convariance matrix.