Research On Sparse Representation Super-Resolution Restoration Algorithm Based On Sample Clustering

The image super-resolution(SR) reconstruction is a resolution enhancement technology which recovers the high-resolution image from one or more low-resolution images with same scenes. SR technology has an important application in some fields such as video surveillance,remote sensing image processing, medical image processing etc. Several sparse representation based algorithms proposed in recent years are introduced and analyzed in this paper, and two novel SR algorithms based on sample clustering are proposed.In SCSR(Sparse Coding Sparse Representation) algorithm, the universal overcompleted dictionary cannot be adapted to variety types of images. To overcome this shortcoming,a SR algorithm based on MCA(Morphological Component Analysis) decomposition and sample clustering is proposed. Firstly, the training feature patchs are clustered by K-means algorithm,and then each clustering is trained to get dictionaries, which are used to process variety types of images. Secondly, the image is decomposed into texture component and smooth component by MCA method. The texture component is reconstructed sparsely and the smooth component is enlarged by Bicubic algorithm. Finally, compared with other SR methods, this algorithm can restore the better image edge details.Moreover, the NCSR(Nonlocal Centralized Sparse Representation) algorithm is proposed to improve the CSR(Centralized Sparse Representation)algorithm in this paper. Firstly, to make the samples to be clustered more stably and accurately, the initial clustering centers of K-means are determined by the distribution of samples during dictionary training. Secondly, the NCSR algorithm model considers the priori knowledge of image by adding the image nonlocal self-similarity constraint, and as a result the reconstruction error is reduced. The experiment result shows that, the proposed NCSR outperforms other SR methods in image quality and the speed of reconstruction.