Research On Sparse Representation Based Image Super-Resolution Reconstruction Method

With the development of technology and industry, many kinds of imaging devices are designed(such as i Phone, i Pad and digital camera, etc). Meanwhile, image has become one of the most important carriers that help us to convey information. However, due to the influence of the disadvantage factors(such as imaging devices limitations and unpredictable external environment, etc), the images captured by the imaging devices are often degraded or low-quality when there exist some processes(such as imaging, transmission, conversion, storage, replication and display, etc). Image degradation means that the useful information is contaminated and lost during the processes described above. To reconstruct them, one is to try to promote and improve the imaging device hardware. Until now, due to the limitation of the manufacturing technology, there have few desired outcomes. The other is to develop the software technology, namely, the current image super resolution technology based on the digital signal processing theory. In recent years, more and more image super resolution theories are proposed and are widely used in almost all of image processing applications in the academic field. In the actual application, the image super resolution technology has been applied in a lot of national key scientific and technological projects. Therefore, the image super resolution technology is significant both in theory and in practical engineering.This paper mainly focuses on the ill-posed image super resolution. Firstly, the present development in this field is reviewed at home and abroad. And then three drawbacks that there exist in widely researched image super resolution technology based on the sparse representation theory are pointed out. Finally, to solve three drawbacks, three image super resolution technologies are developed. The main original innovations of this paper are summarized as follows:â‘ The dual-sparsity regularized sparse representation model is proposed and applied to the image super resolution reconstruction field. In term of the sparse representation model, the accurate sparse representation coefficient for image reconstruction is crucial. Although the existing sparse representation models consider both the local sparsity prior and the nonlocal self-similarity prior to regularize the sparse representation coefficient, they fail to consider the relationship among all entries of the sparse representation coefficient. Hence the modeling capability may be limited. To solve the drawback, firstly, the row nonlocal self-similarity prior of the sparse representation coefficient is explored. Secondly, using this prior, a row nonlocal similarity regularization term with l1-norm constraint is further proposed. Thirdly, this proposed regularization term is introduced to the conventional local sparsity and column nonlocal self-similarity sparse representation model to form the proposed model. Finally, the standard iterative shrinkage algorithm is used to solve the proposed model.â‘¡The low-rank constraint and nonlocal self-similarity based sparse representation mode is proposed and applied to the image super resolution reconstruction field. In the innovative work â‘ , the performance of sparse representation is effectively improved by introducing the row non local self similarity prior. But, due to the independent coding process of each image patch to be encoded, the global similarity information among all similar image patches is lost when they code the patches in whole image. As a result, similar image patches may be encoded as totally different code coefficients. The stability of the sparse representation models may be harmed. One way to enhance the stability is to concern the global similarity information among all similar image patches when they code the patches in whole image. By introducing the low-rank constraint which is better at capturing the global similarity information into the nonlocal self-similarity sparse representation model, the similar image patches can be encoded as totally similar code coefficients in whole image. The linearized alternating direction method with adaptive penalty is introduced to effectively solve the proposed model. Extensive experiments demonstrate that the proposed model has the capable of preserving the global similarity information.â‘¢The image super resolution reconstruction based on the global non-zero gradient penalty and non-local Laplacian sparse coding is proposed and applied to image super resolution reconstruction field. From the point of view to solve the sparse representation coefficients, the innovative works â‘  and â‘¡solve the performance and stability of the sparse representation model. But, due to the limitation of dictionary in learning the image structure, the sparse representation model tends to reconstruct incorrectly the edge structure. To overcome the problem, firsltly, the proposed strategy assumes that the high resolution image consists of two components: the edge component image and the texture detail component image. The edge component image only contains the smooth region component and the major edge component. The texture detail component image only contains the smooth region component, the texture component and the detail component. Secondly, The global non-zero gradient penalty model, which can globally sharpen major edges and preserve their geometric structure by increasing the steepness of transition in a sparsity-control manner, is proposed to reconstruct correctly the edge component image; The non-local Laplacian sparse coding, which can absorb the useful information and exclude unnecessary information, is proposed to reconstruct the texture detail component image. Thirdly, the reconstructed edge component image and the reconstructed texture detail component image are composed to obtain an initial high resolution image. Finally, the global and local optimization model is applied on the initial high resolution image for removing the possible artifacts.Extensive experiments demonstrate that the proposed three image super resolution technologys achieve much better results than many state-of-the-art technologys in terms of peak signal to noise ratio(PSNR), structural similarity(SSIM) and visual perception.