Low-Rank Modeling And External Prior For Single Images Super-resolution

In many practical applications,due to the limitation of image acquisition sensors,the resolution of the collected image cannot meet the application requirements.In order to improve the image resolution,super-resolution(SR)technology is highly required.Among different kinds of methods,the reconstruction-based methods,whose performance mainly depends on the prior information in images,gets much attention.However,most of the SR methods based on reconstruction do not effectively utilize the internal and external prior information of the image,resulting in poor results.In order to solve the above problems,this paper proposes two new super-resolution reconstruction algorithms:(1)In the traditional non-local low-rank(NLR)modeling method,when the fixed rectangular search window is used to find similar image blocks,the surrounding unrelated pixels will be introduced.Therefore,if a group of similar image blocks containing a number of unrelated pixels are arranged by column vectors and combined into a matrix,the generated matrix does not have the property of low rank.In order to solve this problem,this paper proposes a single image SR algorithm based on low rank representation of data adaptation.Firstly,shape adaptive grouping of similar image blocks is carried out by using steering kernel(SK)model.Then,the grouped similar image blocks are used to construct data-adaptive low-rank(DLR)constraint.Finally,an iterative solution algorithm based on split-Bregman algorithm is proposed for the reconstruction problem with DLR constraint.Experimental results show that the proposed algorithm has good SR performance and is superior to many existing algorithms in both subjective and objective quality.(2)In order to further improve the performance of the proposed DLR-SBI algorithm,this paper further proposes to jointly take advantage of the external prior and internal prior information.First of all,the external gradient prior(EGP)is learned by using deep neural networks;Secondly,the SR reconstruction problem with both internal DLR and external EGP constraints is established to more accurately constrain the solution space of the reconstruction problem,so as to obtain high-quality reconstruction results.The experimental results show that the proposed algorithm can obtain the best subjective and objective quality when compared with the existing algorithms.