Research On The Image Super-resolution Reconstruction Based On Spatial Regularization Technique

In the image acquisition process, due to the limitation of imaging conditions andimaging mode, the image resolution acquired usually can not meet the requirementsof practical applications. How to improve the spatial resolution of the image and theimage quality is always the problem to be solved of the image processing technology.The multi-frame image super-resolution reconstruction technique uses the signalprocessing methods to fuse the complementary information between multiple imagesof the same scene so as to obtain a high resolution image. This technology can realizethe observation that is better than the system resolution without changing hardwareconditions, so it is an economic and effective method for improving image resolution.As a result, this technology has a very broad application prospects in many fields andextensive attention by the science and engineering in recent years. Therefore, doingresearch on this topic has a very important theoretical and practical significance.In this thesis, through the system analysis of image super-resolutionreconstruction technique, we carry out the research on the spatial domainregularization based super-resolution reconstruction algorithm as the main clue. Novelimage edge preserving and image denoising models are proposed, which is developedby employing the image prior information and fully considering the human visualeffects on different regions of the image. Then based on these models, series ofmulti-frame image super-resolution reconstruction algorithms that is based on spatialdomain regularization are presented. The main contributions of this thesis include:(1) The high order MRF (Markov Random Fields) model i.e. expert model isintroduced into the super-resolution image reconstruction as the regularization term.Meanwhile, in order to overcome the shortcoming of the conventional expert modelthat will blur the image edge when it is directly applied in super-resolutionreconstruction; a weighted expert field based on spatial information is proposed andused for super-resolution reconstruction. The proposed method uses the curvature difference operator to describe the image spatial structure and defines a weightedfunction to describe the filter that is obtained through the training of the expert field.In this case, the filtering capacity is weakened at the image edge, while the filteringcapacity is enhanced in the flat image region. As a result, suppressing noise andpreserving edge can be realized in the proposed super-resolution reconstructionalgorithm. Simulation and real experimental results show that our method can gethigher peak signal to noise ratio and better visual effect compared to severalconventional super-resolution reconstruction methods.(2) The performance of traditional shock filter in image enhancement is analyzed,and in view of the noise sensitivity problem of it, the gradient vector flow (GVF) isintroduced into super-resolution reconstruction algorithm. Two improved shock filtermodels are proposed, which are formed by separately using the GVF and GVF withimage gradient instead of the image gradient in traditional the shock filter. Thedenoising performance of the improved models is analyzed and its results indicate thatthe two models both have good capacity to remove the image noise, and the combinedmodel has a better effect compared to the direct use of GVF model. Finally, the twoimproved models are incorporated into the image super-resolution reconstructionalgorithm as the regularization term respectively. The Qualitatively and quantitativelycomparative experiments show that the two kinds of modified shock filter models areFeasible and effective for image super-resolution reconstruction.(3) After analyzing the advantages and disadvantages of the total variation(TV) and the fourth order partial differential equation (FPDE) in the aspect of imagedenoising, this paper proposed a combined PDE based super-resolution reconstructionmethod by defining a weighted function that is used to couple the two PDE models. Inthe method, the weight value is set higher for TV model in the edge region in order tosustain the image edge and texture details well, while the weight value is set higherfor FPDE model in the flat area of the image so that the “step effect” produced by TVmodel can be suppressed and the image visual effects can be improved. The combinedPDE model as a regularization term has been used for super-resolution reconstructionof both real and simulated images. The results show that the proposed method can get better reconstruction results due to the combined model involves the advantages ofboth TV and FPDE models.