Research On Algorithm Of Sparse Representation Based Image Super-Resolution

The purpose of image super-resolution algorithms is to gain a High resolution image from one or several images of the same scene. We can obtain more clear person or object that we want to observe by enlarging the target area of the image. Currently, this technology has been widely used in computer vision, HD TV, medical diagnosis, analysis of satellite images and video surveillance. Now there have many image SR algorithms such as frequency domain based, interpolation based, regularization based and learning based. Sparse representation based image SR is a kind of leaning based method and is more advanced algorithm recently.But it also have shortcomings in quality and speed of reconstruction。Firstly, we analysis the efficiency of algorithm which affect the speed of reconstruction.We find that we can improve the efficiency of feature-sign algorithm and thus improve the speed of reconstruction from the experiment results.Then,we improve the speed of reconstruction from two aspect:reduce the number of basis and reduce the size of each basis. We use K-means based feature clustering to reduce the number of basis. In this method,we first cluster the image blocks into different clustering, and then use joint dictionary training method to train corresponding dictionary pair for every cluster. Also, we propose a robust method to determinate the class of the input.Cluster based method can obtain more structured and lower dimension sub dictionaries,thus,we can choose the best relevant sub dictionary for every block in the image reconstruction.This obviously improved the quality and speed of reconstruction from the experiment results.In the aspect of reducing the size of basis,we use PCA to reduce the dimensionality of the basis.This method reduce the dimensionality of dictionary as well as preserve the quality of reconstruction.Lastly, we use adaptive block size to improve the quality of reconstruction.We divide images into three class(Low-complexity, middle-complexity and high-complexity).According to the complexity of the image. In the process of reconstruction, we choose different block size for the image of different complexity(for example, as to upscale size of two, we choose3×3size for high-complexity image,5×5for middle-complexity image and7×7for low-complexity image).In this way, we can guarantee that every block contain appropriate information for reconstruction, thus can improve the quality of reconstruction.The results of our experiments show that improved methods we propose in this paper improved the quality and speed of image reconstruction.