The Image Super-Resolution Based On Sparse Dictionary Learning And Kernel Sparse Representation

In real applications, the most direct and effective method of obtaining the information what we want is image. But in our real researching, image resolutions what are acquired can not often satisfy the requirements, due to the limitations of the image acquisition systems and the effect causing by the environment. These low resolution images not only affect the identification of interesting target, but also reduce the visual perception of images. Thus, in order to guarantee the image visual perception and target identification rates, the super-resolution technique is presented.As an important digital image processing technique, image super-resolution is widely used in the field of security, personal image processing, medical image processing, satellite remote sensing imaging etc. Super-resolution image reconstruction (SR) refers to a resolution enhancement technology that extracting image information from several low-resolution (LR) images. It can eliminate various noise and blur which affect the quality of the image, rebuilding a higher resolution image with more clearly quality. Based on the difference of accessing images, the image super-resolution can be divided into the single-frame image super-resolution and the multi-frame image super-resolution. The multi-frame image super-resolution affects the quality of the image through the complementary information which is provided by many images in the same environment and the similar time phase. In fect, not only the information that can meet to our needs is not enough, but also we can gain the only one image for our research. As the result, it is necessary to learn the single-frame image super-resolution.What’s more, technical difficulty of super resolution reconstruction is that the low resolution image was not sufficient to restore the actual observed high resolution image. Mathematically, the problem is an ill-posed. Super-resolution reconstruction algorithm have many branches. And super-resolution reconstruction algorithm based on sparse is more prominent now. Sparse representation super-resolution was based on theory of compressed sensing data representation model, through the process of establishment of over-completed dictionaries, using samples of high resolution and reconstruction image patch similarity between the sample patch and high resolution patch, to reconstruct of target images.In this article, super resolution reconstruction technique is discussion and analysis. Research is mainly target at single image sparse representation super-resolution technique. The main contribution of this article is as follow:1. This paper describes and analyzes theory and application of Super resolution reconstruction technique. And describes mainstream super-resolution reconstruction technology and analyzes the main problems in the SR reconstruction Finally, discussed the reconstruction of the objective evaluation criteria.2. This paper analyses and discusses a sparse representation of the image and super-resolution reconstruction algorithm based on sparse representation. And for laser remote sensing image super-resolution problem, single image Super-Resolution algorithm based on sparse dictionary learning and kernel sparse representation is proposed. Firstly, a set of many samples is obtained by pre-processing applied on the given high-resolution images. Then low-resolution dictionary and high-resolution dictionary are learned by sparse dictionary learning technology. Finally, we propose Kernel Orthogonal Matching Pursuit technology (KOMP), by which kernel sparse representation problem is solved to obtain sparse coefficients, and super-resolution Image is recovered by these coefficients.Compared with state-of-the-art methods, for using sparse dictionary learning technology, dictionary learning need fewer examples and computation cost, while kernel sparse representation is able to capture the details of image, the super-resolution performance of the proposed algorithm is improved. Our experimental results validate the effectiveness of the proposed algorithm.