Robust super-resolution algorithm for low-quality surveillance face images

Human understanding with image semantic information, especially structural information, is robust to the degraded pixel values. In order to enhance the robustness of traditional methods to low-quality surveillance images, we propose a face super-resolution approach using shape semantic model. This method describes the facial shape as a series of fiducial points on facial image. And shape semantic information of input image is obtained manually. Then a shape semantic regularization is added to the original objective function. According to the correlation of coefficients of image and shape, the variables of reconstruction fidelity term and shape regularization item are unified. And the steepest descent method is used to obtain the unified coefficient. Experimental results of simulation and real images indicate that the proposed method outperforms the traditional schemes significantly both in subjective and objective qualities.