A Non-Linear Warping Method for Face Hallucination Based-on Subdivision Mesh

Image resolution is an important factor affecting face recognition by human and computer. We propose a new nonlinear warping method to align the facial images of the training set, which is then used to predict high resolution facial image from a signal input low resolution facial image. Using sparse initial mesh specified by the facial feature points and a special designed subdivision mesh algorithm, we can refine the initial mesh to arbitrary precision. Hence, we can warp the training facial images to a unified form pixel by pixel. The input low-resolution image is also be warped to the unified form by same way. Then we can use principal component analysis (PCA) to fit the warped input facial image to the linear combination of the warped facial images of the training set. Because the presented warping method can warp all kinds of facial image such as different race or pose to a unified form smoothly, we get well image registration results and hallucinating faces by only very simple PCA global approach. Experiments indicate our nonlinear warping method can enhance the similarity between the input facial image and training facial images, which is the foundation of hallucinating face method. Therefore, our nonlinear warping method can also be used as a pre-align step to improve the performance of the other hallucinating face methods.