| 摘要: |
| 运动模糊核的准确估计是实现单幅运动模糊图像盲复原成功的关键.但是，因为不能准确提取出有利的图像边缘以及简单的正则化约束项的设计，导致现有运动模糊核（motion blur kernel，简称MBK）的估计并不十分准确，存在瑕疵.因此，为了能够估计出准确的运动模糊核，提出了一种基于空间尺度信息的运动模糊核估计方法.首先，为了准确地提取有利的图像边缘，移除有害的图像结构，提出了一种基于图像空间尺度信息的图像平滑模型，实现有利图像边缘的准确快速提取；然后，从运动模糊核的内在特性出发，将空间域的L0范数和梯度域的L2范数结合到一起，提出了一种正则化约束模型，很好地保证了运动模糊核的稀疏平滑特性，并结合之前提取出的有利的图像边缘，共同实现运动模糊核的准确估计；最后，采用一种半二次性分裂的交互式最优化策略对提出的模型进行最优化求解.在客观的评价指标和主观的视觉效果上进行了大量实验，其结果证明所提出的方法能够估计出更准确的MBK和复原出更高质量的去模糊图像. |
| 关键词: 运动模糊图像盲复原 运动模糊核 有利的图像边缘 空间尺度信息 多正则化约束模型 |
| DOI：10.13328/j.cnki.jos.005587 |
| 分类号:TP391 |
| 基金项目:国家自然科学基金（61601070，61271259）；重庆市教委科学技术研究计划（KJZD-K201800603，KJZD-M201900602）；重庆市基础与前沿研究计划（CSTC2016jcyjA0455，CSTC2014kjrc-qnrc40002）；重庆市教委科学技术研究项目（KJ1600411，KJ14004 29） |
|
| Spatial-scale-information Method for Motion Blur Kernel Estimation |
|
TANG Shu, WAN Sheng-Dao, YANG Shu-Li, XIE Xian-Zhong, XIA Ming, ZHANG Xu
|
|
Chongqing Key Laboratory of Computer Network and Communications Technology(Chongqing University of Posts and Telecommunications), Chongqing 400065, China
|
| Abstract: |
| The accurate motion blur kernel (MBK) estimation is the key for the success of the single-image blind motion deblurring. Nevertheless, because of the imperfect useful edges selection and the simple regularizers design, the MBKs estimated by existing methods are inaccurate and contain various flaws. Therefore, in order to estimate an accurate MBK, in this study, a spatial-scale-information method for MBK estimation is proposed. First, in order to accurately extract the useful image edges and remove the pernicious image structures, an image smoothing model based on the spatial scale information is proposed, such that the useful image edges can be extracted accurately and quickly. Then, according to the characteristics of the MBK, a regularization constraint model, which combines spatial L0 norm with gradient L2 norm, is proposed for preserving the continuity and the sparsity of the MBK well. By combining the extracted useful image edges and the proposed regularization constraint model, the accurate MBK estimation can be achieved. Finally, a half-quadratic splitting alternating optimization strategy is employed to solve the proposed model. Extensive experiments results demonstrate that the proposed method can estimate more accurate MBKs and obtain higher quality deblurring images in terms of both quantitative metrics and subjective vision. |
| Key words: blind motion deblurring motion blur kernel the useful image edges spatial scale information multi-regularization constraint model |
