The Research Of Image Restoration Algorithm Based On Regularization Method

Image detection is an important means to collect evidence and extract clues in the process of public security organs. Along with the development of the safe city construction and the expanding scale of the video surveillance network, the mass of video images and other information resources are easy to obtain. However, in the process of capturing and transmitting, the image is influenced by equipment, environment and man-made. The image quality is uneven, and the blurred image occupies a large proportion of the image and video data. The police cannot identify the details of the incident scene from the blurred video images, such as the suspect’s facial features, the license number of accident vehicle. The judgment of investigators is affected. The blurred video image cannot become the scene forensics data of court. Restoring the clear image from the blurred image to assist the public security personnel judgments to settle a lawsuit, and service in public security prevention and control, case detection, intelligence and judgment of police work becomes an important problem which is urgently needed to be solved in the field of public security research. Therefore, this paper carries out a thorough research on the various forms of blurred image restoration. The problem is solved by applying theoretical analysis, model building, algorithm design and experimental verification. The major work and innovation points are listed as follows:Firstly, the denoising method of additive noise pollution image is researched. To address the problem of noise interference in the process of image acquisition, transmission and storage, two denoising algorithms which combine with the estimation of the transform domain and spatial smoothing are proposed. Based on the advantages of the Pyramidal Dual-Tree Directional Filter Bank(PDTDFB), such as directional selectivity, low redundancy, flexible and complete reconstruction, and the correlation between the PDTDFB coefficients, two PDTDFB domain image denoising models are proposed: PDTDFB domain Gaussian Scale Mixture model based on Bayesian least square method and PDTDFB domain multivariate shrinkage model based on maximum a posteriori estimation. In the spatial domain, the non-local means filter is further employed to remove the scratch artifacts after PDTDFB domain denoising to obtain the higher visual effect and objective evaluation. Experimental results show that the two proposed methods are effective.Secondly, nonblind image restoration method of the known point spread function is researched. An image restoration model associated with the shearlet-based sparsity and the weighted anisotropic total variation(WATV) in spatial-transform domain is proposed under the known space shift invariant point spread function. The WATV can enhance the protection of the image edges, while overcoming the staircase effects of the traditional total variation. The sparse representation of the image in the shearlet transform can best describe the texture and detail features. Under the regularization-based framework, a new non-convex objective function using the joint model is formed. To solve this function, a split Bregman-based multivariable minimization iterative scheme is proposed. The original objective function optimization problem is decoupled into several subproblems, which can be solved alternatively. And each subproblem has a closed solution. The restoration experiments of three kinds of images which are degraded by blur and noise verify the effectiveness and robustness of the proposed method.Thirdly, the blind restoration method of global motion blurred image is researched. A two steps image blind restoration algorithm based on 0 norm multi-regularization is proposed. The point spread function estimation stage, the advantages of the 0 norm which can preserve the large edge and suppress the small edge are using to depict the sparse characteristics of the natural image gradient. Then, an adaptive edge selection weight function is added to the image gradient constraint to strengthen the preserving of large edge. In order to guarantee the continuous smoothness and sparsity of the point spread function, and reduce the influence of the noise in the point spread function, a point spread function constraint regularization term with 0 norm and 2 norm is formed. By embedding the multiple constraints into the regularization framework, a point spread function estimation model based on 0 norm is proposed. Under the alternating minimization framework, the proposed model is solved using the split Bregman algorithm and the half quadratic splitting rule to estimate the accurate point spread function. The image restoration stage, a nonblind restoration method with ringing suppression is proposed.Fourthly, the blind restoration method of local motion blurred image is researched. For the local motion blur which is caused by the fixed imaging equipment and moved object, this paper proposes a local blurred image restoration method based on blurred region automatic detection, blurred region segmentation, image synthesis and inpainting. Firstly, the local blurred image is divided into image patches and the singular value decomposition blurred index is calculated. The blurred region and clear region of the image are classified according to the blurred index. Next, the alpha channel map is calculated using Closed-form matting algorithm. The original image is divided into two images by alpha channel map. One image only contains the blurred region, and the other only contains the clear region. Then, the image which only contains the blurred region is restored by the global motion blurred image blind restoration algorithm proposed in the previous chapter. The final restoration image is obtained by compositing and inpainting with the restored region and clear region.Finally, to solve the practical problems by using the proposed algorithms in the paper, plenty of blurred images are collected in the natural environment and the restoration experiments are performed. Experimental results prove the effectiveness and practicability of the proposed algorithms, which provides the theoretical basis and technical support for the image restoration of public security organs’ image detection work.