Position-Patch Based Face Hallucination Using Super-Pixel Segmentation and Group Lasso

Abstract

Traditional super-resolution algorithms utilized samples priors to guide image reconstruction by image-patch. All of them use square or rectangle patch for acquiring prior information. However, fixed size patches will diminish structural information obtained by patches. To make patches gain more structural information, we make two adjustments to the face hallucination: superpixel segmentation and Group Lasso. With super-pixel segmentation, we exploit structural features of human faces by segmenting face images into adaptive patches based on their appearances. Group Lasso provides additional structural information through group selection. Our experimental results show that the extra structural information attained by adjustments has a positive impact on the final reconstructed image.