Image Super-resolution Reconstruction Based On Sparse Representation And Dictionary Training

The learning-based super-resolution reconstruction algorithm has become a popularresearch fields in image processing in recent years. In this paper, on the basis of latestresearch, we perform the research by using the dictionary training and the sparserepresentation theory for single image super-resolution reconstruction.Firstly, we introduce the classification of the image super-resolution reconstructionalgorithms, development of domestic and foreign, the image quality assessment standards.The detail of the sparse representation theory and the traditional learning-basedsuper-resolution reconstruction algorithm is described.Secondly, considering that the traditional super-resolution reconstruction algorithmonly use single joint dictionary, which cannot reconstruct the image informationeffectively. We introduce a coupled-dictionary by using the K-SVD algorithm to train thedictionary. The geometry structure information of the training sample images areeffectively utilized.Thirdly, considering that the contents can vary significantly across different imagesor different patches in a single image，we design the algorithm which use sparserepresentation and the sub-dictionaries to reconstruct image. The principal componentanalysis (PCA) technique and high-quality example image patches are used to learn thesub-dictionaries. The image nonlocal self-similarity is introduced as another regularizationterm.Finally, for fully using the image self-similarity in super-resolution algorithm, wepropose a novel image sparse representation method which exploits adaptively the localand nonlocal redundancies of the image. A nonlocal autoregressive model is proposed andtaken as the data fidelity term in sparse representation. The principal component analysis(PCA) technique and high-quality example image patches are used to learn thesub-dictionaries. We update the sub-dictionaries in several iterations to reducecomputational cost. At the same time, the sub-dictionaries are selected in sparse domainadaptively. Extensive experiments on single image validate that the proposed method, compared with several other state-of-the-art learning based algorithms, achieves muchbetter results in terms of both peak signal-to-noise ratio(PSNR) and structuralsimilarity(SSIM) and subjective visual perception.