Super-resolution Reconstruction Based On Dictionary Training And Sparse Representation

Super-resolution reconstruction, which is purpose to restore lost information in theimage processing and improve the image quality, has been studied widely and become animportant direction in the field of modern image researching. The learning-basedsuper-resolution algorithm has become a hot research direction in recent years. On thebasis of latest research, this paper performs the research on single image super-resolutionreconstruction by sparse representation and dictionary learning, mainly in the followingaspects:First of all, considering the quality of image decreases as magnification increases, wedesign the super-resolution algorithm based on image self-information to overcome theshortage of conventional algorithms which have not make full use of the information fromthe image itself. The algorithm train a dictionary by exploiting the information of imageitself, then reconstruct high resolution image gradationally to decrease each magnification.The experimental results show that, compared with the conventional algorithms based onsparse representation, the proposed algorithm can reconstruct more details.Secondly, conventional super-resolution algorithms train one couple of dictionariesand ignore the relationship between the high and low resolution of the sparse coefficients,so the image patches cannot be sparse represented accurately. To overcome this shortage,we designed the super-resolution algorithm based on multi-dictionary and mappingbetween the sparse coefficients. The algorithm train different dictionaries according to theimage patches with different structure, and solve the mapping between the two sparsecoefficients in each group at the same time. The experimental results show that, comparewith the algorithms based on single dictionary or non-mapping, the proposed algorithmcan produce sharper edges, the quality has improved in terms of both visual perceptionand the peak signal-to-noise ratio.Finally, to overcome the shortage of existing algorithms which explore the priorknowledge from single source, we propose the super-resolution algorithm based on theimage self-similarity and dictionary training, combining the information of the image itself and natural images. In the proposed algorithm, image self-similarity across differentscales is exploited，the reconstruction processed only use the input image in the first；then，natural images are used for training the dictionary and reconstructing the initial resultwhich is regarded as the input one. The experimental results show that, compare with theother algorithms, the proposed algorithm can reconstruct more details and improve thequality in terms of both visual perception and the peak signal-to-noise ratio.