With the development of face recognition techniques, illumination problem becomes a main obstacle. It is well-known that in most cases image variation due to lighting direction changes is more than that due to different identities. In this paper, we focus on lighting problem under the scenario of face recognition and the major works in this thesis are as follows. Face representation under various lighting conditions Excluding other factors, such as pose and expression, quotient image, illumination cone and spherical harmonic methods explain the image variations due to lighting changes from different viewpoints under different assumptions. Illumination alignment The basic idea of illumination alignment is illumination normalization, which converts images from different lighting conditions into given one. As we can not control the lighting condition of input image, we usually convert the lighting condition of face database into that of the input image. Photometric stereo based 3D face reconstruction It is an ill-posed problem to recover 3D shape from image. By acquiring more than three images, we design a hardware system to reconstruct face 3D shape. Self-quotient image The most interesting characteristic of self-quotient image is that it needs not any training set and alignment. Moreover it has de-shadow ability. Application in real system According to our experiments on real system, the two algorithms, illumination alignment and self-quotient image, can significantly improve the system recognition rate.
修改评论