局部加权全变差下的盲去模糊  

Blind Deblurring Based on Local Weighted Total Variation

作　　者：吴晓旭[1] 刘秀平[1] 刘宇[1] 曹俊杰[1] 王军 苏昕[3] 

机构地区：[1]大连理工大学数学科学学院,大连116023 [2]鞍钢股份-大船重工大连钢材加工配送有限公司,大连116021 [3]南开大学物理科学学院,天津300071

出　　处：《计算机辅助设计与图形学学报》2014年第12期2173-2181,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61173102)

摘　　要：图像去模糊是图像处理和分析中的基本问题之一,其本身是一个不适定问题,通常需要使用正则化方法来提高求解过程的稳定性.为了解决去运动模糊问题,从图像的局部特性出发,提出一种基于局部加权全变差(LWTV)的正则化方法,并给出了一种基于交替迭代的有效解法.针对非盲去卷积问题,为了克服传统全变差(TV)正则化方法的不足,以图像局部的变化信息为权值,在加大对图像中平坦区域的惩罚力度的同时,减小对图像中边缘区域的惩罚力度;针对模糊核估计问题,首先利用相对全变差(RTV)方法提取图像的显著性结构,然后利用显著性结构进行初步模糊核估计,再采用LWTV模型进行临时清晰图像估计,通过以上3步交替迭代获得最终的模糊核.实验结果表明,该方法可以在去除模糊及噪声的同时,很好地保持图像边缘并抑制振铃效应.Image deblurring is one of basic problems in the field of image processing and analysis. Since it is an ill-posed problem, a regularization is required to improve the stability of the solving process. In this paper, we propose a regularization method for motion deblurring based on local weighted total variation （LWTV） in terms of the local features of the image, and its corresponding solution based on alternating minimization. In the part of non-blind deconvolution, to overcome the shortcomings of traditional total variation （TV） method, we adopt the local variation of image as weights to increase the punishment on the flat area and reduce the punishment on the edge area. In the part of kernel estimation, we first extract the significant structures using relative total variation （RTV） method, then estimate initial kernel with the significant structure, and finally estimate the temporary image using LWTV model. In this way, the kernel can be obtained by alternating above three steps iteratively. Experimental results show that the proposed deblurring method can not only remove the blur and noise, but also keep the sharp edge and suppress ringing artifacts.

关 键 词：图像去模糊 局部加权 全变差 模糊核估计 盲去卷积 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]