The appearance of human face is determined by the interaction between light and matter on the face. Changes in illumination conditions or surface reflectance properties of human face can lead to significant variations in the appearance of face. Therefore For a long time face image analysis under variable i11umination is an important task to computer vision society. At first this thesis gives a brief retrospect to the related work. Then we focus on the following two subproblems: face de-lighting & relighting, and face alignment under variable illumination. It is very difficult to recover the illumination-invariant reflectance from a single input image because it is an ill conditioned problem. In our first work structure similarity between the reflectance image and its illuminated image is revealed firstly based on an ideal class definition. Then a simple linear reconstruction scheme is successfully proposed for reflectance rendering, in which the reflectance face images are learnt (or generated) from the training examples in terms of this structure similarity. Principal Component Analysis is adopted to get a robust rendering result. Extensive experiments show that the proposed method is efficacious in separating reflectance from an input face image, and reducing the variation caused by lighting conditions. Furthermore, this method can be used directly to render new images of the input face under different lighting conditions. Our second work presents an approach to face alignment under variable illumination, an obstacle largely ignored in previous 2D alignment work. Firstly we introduce what is face alignment and why the traditional methods to face alignment can not handle illumination variation. Then we discuss how to employ two forms of relatively lighting-invariant information to account for illumination variation. Edge phase congruency is adopted to coarsely localize facial features, since prominent feature edges can be robustly located in the presence of shading and shadows. To accurately deal with features with less pronounced edges, final alignment is then computed from intrinsic gray-level information recovered using a proposed form of local intensity normalization. With this approach, our face alignment system works efficiently and effectively under a wide range of illumination conditions, as evidenced by extensive experimentation. At last we discuss some possible extensions of our research work and point out their perspective applications.
修改评论