Research On Single Image Super Resolution Algorithm Based On Self-similarity

In recent years super-resolution algorithm based on self-similarity produces good visual results, which shows good prospects for development. So this dissertation focuses on the study of super-resolution algorithm based on self-similarity.Firstly, by analyzing the advantages and disadvantages superresolution algorithm based on self-similarity, the dissertation proposes a super resolution algorithm considering self-similarity and steering kernel regression. The algorithm uses the gaussian pyramids to find the global similarity patch for image construction, and uses the kernel regression for local structure constraints. The dissertation first introduces the kernel regression model. Then the algorithm based self-similarity with steering kernel regression is proposed. Finally we verify the performance of the proposed algorithm.Moreover, LSE(Local Self-Examples) algorithm still has room for improvement in visual results. The dissertation uses the traditional pyramid similarity hypothesis to propose an improved LSE algorithm. By introducing horizontal pyramids and vertical pyramids to extend the traditional gaussian pyramids, the given method uses the corresponding position patches in the improved pyramids. The algorithm do not need global search to find the best patch, and only need calculate the corresponding position patches for finding the best match patch. Finally we verify the performance of the proposed algorithm.Last, super resolution is sensitive to noise, the dissertation proposes a super resolution algorithm for noisy image. By combining the de-noising and super-resolution, we propose noisy image super resolution algorithm based on improved pyramids. The dissertation firstly analyzes the noisy image similarity using the improved pyramids, then gives the noisy image super resolution algorithm based on improved pyramids. Finally, we verify the results about noisy image scaling.