Research On Sparse Representation Based Image Super-resolution Reconstruction Model

With the rapid development of the information age,digital images have gradually become the most important carrier for people to transmit information due to their excellent characteristics,and thus they are widely applied to our daily applications.In terms of digital images,spatial resolution is an important indicator of its quality.However,during the actual imaging process,the resulting images often tend to have low resolution due to the physical limitations of the optical imaging systems and various degradation factors,which brings many difficulties to the subsequent image processing and image analysis steps and is not conducive to accurately understanding the objective information contained in the images.Therefore,it is very important to find a way to improve the image resolution effectively.Consequently,image super-resolution reconstruction technology is the effective means to solve the above problem.This technology is able to transcend the inherent limitations of existing imaging systems,and accordingly enhance the resolution of low resolution images via image processing.For this reason,super-resolution reconstruction technology has received great attention from the international academic and business community,and has become one of the most appealing research areas in the field of image processing.Recently,sparse representation,as a new and effective image model,has injected a new vitality into solving the single image super-resolution(SISR)reconstruction problem,and has also been a hotspot issue of SISR.Faced with this new opportunity and challenge,this dissertation concentrates on the study of the highly concerned sparse-representation-based SISR model,and proposes several schemes for improving the deficiencies in the existing model.The main contributions of this dissertation can be summarized as follows:(1)The analysis of the existing sparse representation based SISR model.First,the single image degradation model is established,and the various degradation factors are analyzed so as to deduce the ill-posedness of SISR problem.Subsequently,the theory of signal sparse representation and the existing sparse representation based SISR model is introduced.Finally,we analyze the aforemetioned model in-depth when it is dealing with the SISR problem,and accordingly point out its shortcomings in the following aspects: dictionary learning,sparse coding,and model efficiency,providing the theoretical basis for the following parts.(2)Dictionary learning is known as a key procedure of sparse-representation-based SISR model.To overcome the shortcomings of the currently used sparse dictionary,a dictionary learning and adaptive selection method is proposed based on geometric features of image patches.However,unlike the conventional way that learns a universal and over-complete dictionary to code different varieties of local structures,the combination method of local orientation estimation-based image patch clustering and principal component analysis is used to obtain a series of geometric dictionaries of different orientations during the dictionary learning process.Subsequently,only the dictionary of the nearest orientation is adaptively assigned to each of the input patches that need to be represented in the sparse coding process.Moreover,the consistency of gradients is further incorporated into the basic framework to make more substantial progress in preserving more fine edges and producing sharper results.Experiment results indicate that our proposed method outperforms many similary counterparts in terms of both numerical indicators and visual quality.(3)Resulting accurate enough coefficients during the sparse coding process is the premise and basis for the sparse-representation-based model to correctly solve the SISR problem.However,due to the complexity of the image degradation,at present it still remains a challenging work to recover the ideal coefficients from the severely degraded data as accurately as possible.For this reason,a research strategy is proposed based on sparse coding error compensation.First,a comprehensive analysis of sparse coefficient noise is conducted to verify the necessity of introducing sparse coding error compensation.Then,inspired by the self-similarity and multiscale redundant information of natrual images,a multiscale redundancy and sparse representation based SISR model is proposed.This model is able to compensate the sparse coding error by using the redundant information in the multiscale image space during the sparse coding process,thereby improving the accuracy of the sparse coding coefficients.In addition,considering the high computational complexity of searching information in the multiscale space,anther improved SISR method is further presented based on the proposed bidirectionally aligned sparse representation(BASR)model.In this improved model,the bidirectional similarities are first modeled and constructed instead to form a complementary pair of regularization terms.Then the raw sparse coefficients are additionally aligned to this pair of standards to restrain sparse coding noise and therefore result in better recoveries.Extensive experimental results demonstrate that the proposed procedures of sparse coding error compensation can improve the model performance and the robustness to various degradation factors.(4)In view of the low efficiency caused by the inherent deficiencies of conventional sparse representation theory,a model optimization strategy is developed by the simultaneous use of deep learning and sparse representation.According to the in-depth analysis of the existing residual-learning-based SISR model,an improved network structure is presented by taking advantage of the proposed component learning.The core idea and difference of this learning strategy is to use the residual extracted from the input to predict its counterpart in the corresponding output.To this end,a global decomposition procedure is designed on the basis of convolutional sparse coding and performed on the input for extracting the low-resolution(LR)residual component from it.Owing to the good properties of this decomposition,the represented residual component still stays in the LR space so that the subsequent part is capable of operating it economically in terms of computational complexity.Thorough experimental results demonstrate the merit and effectiveness of the proposed component learning strategy,and our trained model outperforms many state-of-the-art methods in terms of both speed and reconstruction quality.