Research On Image Reconstruction Based On Group Sparse Representation And Non-local Similarity

As the development of mutimedia communication and information technology, image processing techniques are becoming significantly more inevitable, with its broader applications worldwide. Among the rigorous research in the areas, image super-resolution reconstruction and image denoising both receive much attentions and occupies irreplaceable positions in daily life and industrious productions. This paper did research on the two orientations and discuss the super-reoslution algorithm and image denoising algorithm. As the image modeling and representation are the fundamentals in image processing area, the qualify of the model determines the effectiveness of the following representation and post-processing. Nowadays image modeling based on sparse representation has turned into the central issues-and is widely emphasized by researchers worldwide.However, the amelioration of hardware technology may not be achieved in a short term due to the involving high cost and technical difficulty. To overcome this limitation, researched has proposed some other methods to improve the resolution of image to provide people with more details, without changing the hardware device. Image super-resolution reconstruction is characterized for constructing images with high-resolution through low-resolution degraded images. This method provides a solution to the shortage of hardware by incorporating signal processing theories, and has become a spotlight in the many research areas like image processing and computer visualization.As a newly developed image representation model, sparse representation is able to represent any image by using the liner combination of a few elements through appropriate over-complete dictionary. Nowadays spares representation has attracted much attention worldwide, and image super-resolution reconstruction based on sparse representation has also aroused great interest in the research areas. However, the traditional method assumes that the sparse representation coefficients are randomly distributed, failing to consider the special internal structure of sparse signals. Research displays that sparse coefficient tends to exhibit a special group or tree structure. The introduction of a sparse coefficient group structure can enhance the performance of sparse representation, which is now being widely studied by researchers. In this dissertation, we systematically analyzed the applications of dictionary training and sparse coding algorithm in the area of image super-resolution reconstruction. The main content includes three aspects:1. The basic concepts, mathematical models and optimized algorisms involved in sparse representation theory are described.2. To solve the issue of sparse super-resolution reconstruction, we take into account the internal structures of sparse representation and adjust them with grouping methods. By combining non-local correlations in images, we propose a group structure based on sparse and non-local self-similar image super-resolution reconstruction algorithm to improve the quality of reconstructed image.3. Image denoising by using the traditional K-SVD dictionary training algorithm. As the-sparse doesn’t consider the internal structures of its coefficients, we group these coefficients to achieve an improved over-complete sparse representation. To better utilize the relations among images, we proposed a model image denoising model based on group structure sparse and nonlocal self-simlarity.