Robust Face Recognition Methods under Illumination Variations toward Hardware Implementation on 3DCSS

1. Introduction In order to realize highly intelligent information processing systems which can recognize various objects in real-time/real-world, we have been proposed a concept of multi-object recognition system based on 3D custom stack system (3DCSS) presented in the 21st century COE of Hiroshima University [1]. Recently, we have developed a real time human face detection/recognition software system based on eigenfaces methods and implemented it on FPGA [2, 3]. However, there is a problem that the recognition performance of our developed system deteriorates under illumination variations. These variations can not be avoidable in the real world. In this paper, we propose robust face recognition methods which overcome illumination problem by applying three image processings to input images. We show the effectiveness of proposed methods by numerical simulations. 2. Preprocessing for removing the influence of illumination variations The Yale Face Database B[4] is a free database for illumination and pose problems. This database contains 5760 single light source images of 10 individuals each seen under 576 viewing conditions (9 poses x 64 illumination conditions). For all the sets in the frontal pose, the coordinates of the left eye, right eye, and mouth in each image have been appended. The examples of frontal images in the database are shown in Fig. 1. 45 images out of 64 is assigned to one of four Subsets according to the light-source directions. The examples of frontal images belonging to each Subset are shown in Fig. 2. As shown in Fig. 2, although all of these images are same individual, the appearance of these images varies strongly according to the light-source directions. We should remove the influence caused by illumination variations in order to achieve accurate face recognition. Therefore , we apply image preprocessing which is combination Subset 1 Subset 2 Subset 3 Subset 4 Figure 2: Example of cropped frontal images of a single individual in the Yale Face Database B under different illuminations. of Histogram equalization, Laplacian of Gaussian filter and Contrast adjustment to appearance-based recognition methods. (1)Histogram equalization Histogram equalization is most widely used method to enhance biased contrast image that some pixels are concentrated on a narrow range of the pixel intensity. Figure 3(a) shows results of histogram equalization for images shown in Fig. 2. Thus, the quality of the image is improved and feature parts such as eyes, nose and mouth are made more clearly. (2)Laplacian of Gaussian filter In order to …