Research On Regularization Image Restoration Method Based On Prior Constraint Models

With the development of the information age, the requirements of scientific researches and practical applications for high quality signals and images are becoming more evident. Thus, advanced digital signal and image processing technologies have been paid much attention on by both academic and commercial fields. Effective signal and image processing technologies will also lay the foundation for subsequent applications. Regularization is an important research topic in digital signal and image processing. Regularization not only has a deep foundation in mathematics, it has a close connection with optimization theory and partial differential equations, but also it can be combined with various signal and image prior models to solve tasks with different application background. Image restoration is an indispensable pre-process procedure in image processing and pattern recognition. The results of image restoration are able to decide whether subsequent image processing tasks can be well completed.This paper mainly focuses on the studies of the image prior models and image restoration methods based on regularization, both theoretic methods and theirs applications. The paper researches on a variety of image prior constraint models and their applications in image restoration, fast solving methods on regularization models of image resotroation, as well as the blind image resotation methods which do not know the blur information, and applied these reaserch to the restorationes of nature images and remote sensing images with multiple degradation factors. This paper combines theoretical research and application, main innovation theory and research results have been proposed as following:1) Analysis traditional image restoration methods for irregular sampling to regular sampling, deblurring and denoising intensively. Point out the shortcomings of these methods: traditional methods need iterate repeatedly, thus the efficiency of the algorithm is low. Moreover, traditional methods can hardly preserve the details of the image, and can induce artificial defects. To solve the above problems, a restoration method of irregular sampled remote sensing image based on non-local total variation is proposed. The proposed algorithm can eliminate irregular sampling effect, deblur and denoise at the same time. The proposed regularization model can be solved by using operator splitting method efficiently. Experiments show that, compare to other traditional approach, the proposed algorithm can reduce the staircase effect effectively and improve the detail information of the restored image.2) In terms of solving the total variation image restoration problem, efficient regularization solving methods in recent years have been studied. Based on the Bregman iteration, a fast iterative total variation image restoration algorithm is proposed. The proposed algorithm split the original total variation problem into sub-problems that are easy to solve using alternating minimization method and can achieve convergence in less iteration, thus simplify the problem and improve the efficiency. Moreover, to better preserve the details, non-local regularization is introduced into the algorithm, and a method to choose the nonlocal parameter point-wise and adaptively is proposed. Experiment results illustrate that the proposed algorithms have fast converge speed and better restoration results compared to other regularization methods.3) Analysis the patch-based image model, image restoration model using sparse representation and dictionary learning methods intensively. Point out the shortcoming of the patch-based image restoration methods using sparse representation:owing to the discontinuity of these models, stitching effects and artificial strips may appear in the restored image especially when the image is seriously degenerated. To solve the above problems, a projection based sparse representation and nonlocal regularization deblurring and denoising image restoration algorithm is proposed. The algorithm combines sparse representation via adaptive learned dictionary and nonlocal total variation, and use the degenerated image itself to learn the adaptive dictionary. The dictionary learning approach is easy to realize and can represent the image well. The proposed regularization model is divided into three projection sub problems to solve to improve efficiency. Experimental results show that the proposed algorithm can preserve the detail information effectively, and have nice restoration results for images with different degree of degradation.4) In terms of blind image restoration based on partial differential equations, analysis P-M model and shock filter based methods intensively. Point out the shortcoming of these methods in common:heavily depend on the calculation of the gradient value and gradient direction, however owning the calculation way is too simple, these methods cannot get the precise value and direction of the gradient, especially in the degenerated images. So, the restored images lose much information of the original image, more obvious in the corners and small objects. To solve these problems, a shock-diffusion model is presented to restore blurred and noisy images. The proposed approach uses a half smoothing kennel to get the precise directions of the edges, and uses different shock-diffusion strategies for different image regions. Experiment results show that the proposed model can effectively eliminate noise, enhance edges and preserve small objects and corners simultaneously. Compared to other methods, results of the proposed method have better visual appearances and qualitative measurements.