An Improved Method Of Face Illumination Compensation Based On Quotient Image

As one kind of pattern recognition, face recognition has been widely applied in information security, door control system. Many factors contributing to its performance, such as light, gesture and face expression which are the three most important factors. Especially, light plays a more significant role in face recognition. A lot of research shows that the similarity coefficient is larger for the different face image under same light condition than the same face image under different condition. This paper focuses on the theory and application analysis for illumination compensation in face recognition.First of all, this paper summarizes the research status about face recognition and light problem from broad view. Based on that, it introduces two methods:SFS and quotient image method. They are both methods based on two-dimension image sample synthesis. Quotient image method is better than SFS on the fact that it requires fewer samples and simple computation. The contribution for this paper is improving and applying quotient image method. Under the analysis of traditional quotient image method in theory and practice, this paper finds some drawbacks in quotient image method, which limits its effect in reality application. For example it has poor performance for the disposing of shadow. In all, the improvements for this paper are as follows:(1) Increasing the numbers of the light source from three to nine so that any complex light condition can be more precise approximated. (2) Replacing average face by the approximation face from PCA method, it can better satisfy the assumption that all shapes of human face are the same. (3) By specifying the condition for illumination compensation, this paper estimates the unnecessary for illumination compensation so as to reduce the calculation. (4) When calculating the quotient image, this paper uses logarithm transformation to change the matrix quotient image calculation to subtract. By doing this, the calculation and the truncation error caused by matrix quotient calculation are reduced. Finally, this paper test the algorithm in Extended Yale B database, the results of the experiments show that the method in this paper can widely reduce the calculation and effectively increase the recognition rate at the same time. When the light condition changes significantly, this paper shows better performance.