Image Super-Resolution Via Patch Similarity Guided Dictionaries Learning

Image super-resolution is one of the research hotspots in computer vision.It can compensate the poor hardware accuracy and present a real scene.The image super-resolution have wide application prospects in security,remote sensing,medical science and HD shows and so on.This paper focuses on the improvement of image super-resolution to spread related research.The learning-based method is improved from two aspects: improving the training efficiency of learning dictionary and improving the quality of reconstructed image.Firstly,aiming at the inefficiency of traditional dictionary training,this paper proposes a new method to judge the similarity of image patches and construct training dissimilar samples.Our method improves the efficiency of dictionary training and the quality of reconstructed image.Secondly,a new method based on Gauss mixture model is proposed to overcome the limitation of the ability of global dictionary.This method has high efficiency,and the obtained dictionary has strong local adaptive expression ability.Our method can effectively improve the quality of the reconstructed image.Main work in the paper is as following:(1)The related concepts and methods of super-resolution are studied,including the concept of super-resolution,image degradation model and several classical super-resolution methods.(2)Sparse representation-based method is emphatically studied.In order to overcome the inefficiency of traditional algorithms in training dictionaries,we proposed a method to judge the similarity between patches.We use this method to construct dissimilar sample and train dictionaries.Our experiments showed that the dissimilar dictionaries can effectively improve the quality of the reconstructed image and reduce the training time.(3)Considering the weakness of global dictionary to express various image patches,This paper introduces Gauss mixture model to train multi-pair learning dictionaries by using patch group.Our experiments showed that the multi-pair local dictionaries is one of the effective ways to improve the quality of super resolution.Lots of experimental results show that the proposed method can effectively remove redundant image patches and improve the efficiency of dictionary training.The proposed multi-pair of dictionaries can effectively improve the quality of super-resolution.The reconstructed high-resolution image has more texture and clear edges.