The Research Of Image Super-resolution Restoration Based On Sparse Representation

With the development of science and technology, people’s demand for information increases with each passing day, and the image has become one of the most important carriers that help us to convey information. However, due to the influence of the unfavorable factors such as the inherent physical limitations of imaging devices and unpredictable external environment, the images acquired by the imaging devices are often degraded. Among numerous image applications, it is very urgent to reconstruct high-resolution image based on the existing imaging devices and the observed low-resolution image. Image super-resolution restoration technique, which utilizes the prior constraint or the representation model of the image, is one of the ways to solve this problem. As a new image representation model, sparse representation can be used to describe the intrinsic structure and essential attribute of the image, which is widely used in the field of image processing, and has achieved good performance. The paper makes an exploratory and innovative study of sparse representation based image restoration.The main contents of the paper include:(1) The research status of image super-resolution restoration is discussed. The theoretical basis of image super-resolution restoration is introduced. The basic concepts,mathematical models and optimization algorithms of sparse representation and its applications in the field of image processing are introduced.(2) Aiming at the problem of asymmetry between training and reconstruction phase in the classical joint dictionary learning super-resolution method, an off-line couple dictionary learning based image super-resolution method is proposed. The goal of our couple dictionary learning is to minimize the sparse recovery error of the high-resolution sample image patches, and the sparse coding is obtained only by low-resolution sample image patches. This dictionary learning strategy ensures that the sparse representation of a low-resolution input image patch can well reconstruct its corresponding high-resolution image patch. In order to establish the sparse association between the high-resolution and low-resolution image patches more effectively, the gradient features of the low-resolution image patches and the correspondinghigh-resolution residual image patches are used as paired samples for dictionary training.In the reconstruction phase, the high-resolution residual image is reconstructed from the low-resolution input image using the learned dictionaries and compensated, which can recover the high-frequency details of the high-resolution image more accurately. The proposed method has a good performance on super-resolution enlargement.(3) The pure learning based super-resolution method has limited super-resolution ability for blurred images. Aiming at the problem, the non-local self-similarity and hyper-laplacian distribution is introduced as regularization constraints into the sparse representation based super-resolution reconstruction framework, which can make use of the specific advantages and complementary characteristics of different priors. In order to make the sparse domain better represent the underlying image, the high-frequency features are extracted from the underlying image patches for sparse representation. The dictionary learning is integrated into the super-resolution process, in such a way the dictionary can be learned directly from the currently recovered high-resolution image feature patches. This dictionary learning method makes full use of the specific structure of the currently recovered image, enables the dictionary to update adaptively with the currently recovered image, as well as reduces the number of training samples and improves the efficiency of dictionary construction. The proposed method can achieve good super-resolution reconstruction for blurred and noisy low-resolution images.(4) Image super-resolution methods generally use mean square error as optimization criterion, this measure doesn’t consider the structure of the image space,which makes the reconstruction results may not have good perceived quality. Aiming at the problem, a sparsity regularized image super-resolution method based on structural similarity is proposed. The structural similarity measure is introduced into the sparse representation to improve the information preserving term, which makes the sparse recovery image patches preserve the sturcture information of the original image patches better. The structural similarity is also introduced to measure the fidelity between the low-resolution image and the high-resolution image, thus constructing a new image super-resolution reconstruction model. In order to further improve the quality of super-resolution image reconstruction, the non-local self-similarity as the regularization constraint is introduced into the model. For the optimization problem of the sparse representation model and the super-resolution reconstruction model understructural similarity index, an effective numerical solution algorithm is presented. The proposed method can achieve good visual quality and effectively preserve image edges and texture structure.