Face Detection And Recognition Under Variant Illumination

Face detection and recognition is an important research subject in the field of pattern recognition and computer vision. Because of its hidden operation, can be accepted by users directly, friendly and easily, low hardware requirements, face recognition technology has been widely used in many areas, such as authentication, video surveillance, identification and home entertainment. Under controlled conditions, face detection and recognition has basically met user’s needs with their cooperation. However, when it comes to uncontrolled conditions, such as variable illumination, pose variations, object shelter and mixture of emotions, the accuracy of face detection and recognition will obviously decline. Only by solving these issues, face detection and recognition technology can truly be used in real-world applications.In this paper, face detection and recognition under variable illumination is studied. This paper mainly focus on the following questions:the false detection in Adaboost algorithm, feature extraction in face recognition and the impact of illumination on face detection and recognition. The major contributions of this paper are summarized as following:(1) In order to reduce the high false face detection rate caused by Adaboost, this paper proposes an algorithm which combines skin verification and Adaboost algorithm. First, Adaboost algorithm is used in the image to obtain the candidate face regions, and then regional skin model is used in the candidate face regions to detect skin. If the rate of the skin exceeds a certain threshold in a candidate face region, this region is identified as human face and marked; else this region is identified as non-face.(2) In order to reduce the impact on skin detection caused by illumination, this paper proposed an improved "reference white" algorithm based on the study of "reference white" algorithm and combined with histogram equalization. The experiments show that, compared with the traditional "reference white" algorithm, the improved "reference white" algorithm has a better treatment effect, and is more conducive to skin detection.(3) This paper proposed a facial feature extraction algorithm which combines weighted PCA and weighted NLDA, that is:WPCA+WNLDA. A matrix is designed for the PCA algorithm, to appropriately undermine the principal components which are sensitive to illumination changes; a matrix is designed for the NLDA algorithm, to give larger weights to the projection directions which have better classification performance. The experiments show that, compared with Eigenface algorithm which is based on PCA and NLDA algorithm which is based on LDA, the algorithm proposed by this paper has a higher recognition rate.(4) Self-quotient image algorithm is an illumination treatment method which is based on illumination invariant feature, commonly used in face recognition. When there are special light areas or special dark areas in the image, treatment effect of the Self-quotient image method is not ideal. To solve this problem, an improved self-quotient image algorithm is proposed, which is called "Single-Light-Area&Single-dark-Area Self-quotient image (SLA&SDA-SQI)". First, special light areas and special dark areas are detected in the image, if exist, they are handled under the premise that does not change the facial feature, which will convert the input image into an image which has single light area and single dark area, and then apply the self-quotient image algorithm. In order to search the special dark areas more accurately, this paper created a special shadow model, based on the "three court five eyes" theory. The experiments show that, compared with Single Scale Retinex algorithm and Self-quotient image algorithm, the algorithm proposed by this paper has a better treatment effort.