Face Image Super-resolution Reconstruction With Compressed Sensing

In the field of digital image processing, image super-resolution is one of the keyIndicators to evaluate the quality of image. Since face image play an important role incomputer vision, pattern recognition, video surveillance and other field, research aboutface super-resolution has been a hot topic.Rely on software method of improving image resolution is an effect way withoutadditional hardware device cost. In recent years, with the proposed of compressedsensing, many researchers have applied compressed sensing into imagesuper-resolution and have achieved better results than traditional methods, it has greatsignificance for image super-resolution.This paper focus on the problem of single face image super-resolution viacompressed sensing. Research on image statistics suggests that image patches can bewell-represented as a sparse linear combination of elements from an appropriatelychosen over-complete dictionary. Inspired by this observation, we seek a fast faceimage hallucination algorithm based on sparse representation, furthermore, we employseveral methods to improve the image quality. Experiment demonstrate our approach isfeasible.In order to improve the accuracy of the reconstruct image, in dictionarylearning stage, we clustering the training samples and in the reconstruction stage weemploy the over-complete dictionary which trained by the training samples similarwith the input image. Our approach is based on image patches, since the imagesuper-resolution with sparse representation work well on patches with visually salientregions, our algorithm auto choose different reconstruction algorithm from imagepatch’s nature. Moreover, we learn a neural network model for fast sparse inference.Extensive comparisons with state-of-the-art super-resolution algorithms validate theeffectiveness of our proposed approach.