Reasearch On Anisotropic Total Variation And Sparse Image Decomposition-based Image Deblurring Method And Its Applications

With the development of the imaging technology,low-power,low-cost and high-performance imaging systems have played more important roles in disaster relief,expedition,target tracking and identification,and battlefield information perception.However,images captured by cameras often face the image degradation phenomena in imaging process due to some unavoidable factors such as,aberration of optical system,the noise of transmission and storage sensors,the shake of the camera or the relative movements between the scene and the sensor.This paper analyzes the structural characteristics of different imaging objects and proposes several image restoration methods to improve the quality of the recovered images.Moreover,robust image restoration methods are also developed to meet the demands of the wide practical applications.In order to improve the effect of the existing methods on edge preservation,especially on the corner of the edges,a non-blind image restoration method is proposed by using anisotropic total variation(ATV)as the regularization of the edges.ATV has the Wulff geometry characteristics and can describe the image edges well.In order to extend the applications of the previously proposed method,the paper further proposes a blind image deblurring model by regularizing both the image and blur kernel with ATV.The proposed methods not only can enhance the edges of the recovered images where the traditional methods are in a weak position,but also can protect the corner of the image edges where total variation-based methods often make distortion.In the non-blind deblurring,the PSNRs(peak signal-to-noise)and the SSIMs(structural similarity index method)of the recovered images are respectively increased by 15%and 5%compared with other non-blind methods.And in the blind deblurring,the PSNRs and the SSIMs have a growth of 20%and 6%,respectively.Moreover,the proposed blind method significantly improves the accuracy of the estimated blur kernel and decreases the loss of the kernel and recovered images.At the same time in the restoration of image edges,this paper also researches how to recover the images full of textures.This type of image is often called texture images which often contain two different parts:cartoon(the structure and flat parts of images)and texture(the oscillating parts in images).Although the image deblurring methods based on total variation(TV)decomposition can improve the quality of the recovered images to a certain extent,the stair-casing effect produced by TV restricts the further improvement of image quality.In order to further improve the quality of texture images,a cartoon-texture decomposition model is proposed and used for non-blind image deblurring.The proposed method respectively regularizes the texture with the sparsity of discrete cosine transform domain,and the cartoon with a combined term including framelet-domain-based sparse prior and a quadratic regularization.For the restoration of the texture images with the blur kernel unknown,this paper further proposes a blind image deblurring method which employs the same regularization terms as before on cartoon and texture,and uses a combined term including TV and a quadratic term to improve the quality of the estimated blur kernel.Both the proposed non-blind and blind methods can simultaneously recover the different contents of images and decompose an image into separate parts under different scales.In the non-blind deblurring,compared with other non-blind methods,the proposed method can decrease the loss of information by about 26%,and raise the PSNRs and SSIMs of the images by 10%and 5%,respectively.In the blind deblurring,the error ratio decreases by 50%,and the PSNRs and SSIMs are separately increased by 8%and 31%.In addition,the proposed blind method can estimate more accurate blue kernel by using image decomposition.Based on the achievements of the edge and texture restoration research,this paper extends the proposed methods into the blur problem solving caused by the high linear velocity in the tilt rotating scanning imaging(TRSI)system.It firstly models the image degradation process as the combination of the rotation and translation,and then ignores the effect of the translation in the final model according to the practical application cases.To effectively recover blurred images captured by TRSI system,this paper extends the previously proposed methods from the space-invariant blur kernel based model into the space-variant kernel model.The results demonstrate that our methods can effectively solve the blur problems in TRSI process.