一种基于结构提取的图像盲去模糊方法  

A blind Image Deblurring Method Based on Structure Extraction

作　　者：袁皓玉 邓晨 YUAN Hao-yu;DENG Chen-quang(The 27th Research Institute of China Electronics Technology Group Corporation,Zhengzhou 450047,China;No.92932 Unit of PLA,Zhanjing 524000,China)

机构地区：[1]中国电子科技集团公司第二十七研究所,郑州450047 [2]中国人民解放军92932部队,湛江524000

出　　处：《电光系统》2024年第3期10-17,共8页Electronic and Electro-optical Systems

摘　　要：针对图像盲去模糊问题,即在模糊核未知情况下由观测到的模糊图像恢复原始的清晰图像,常用的方法是先估计模糊核,然后用估计的模糊核通过非盲去模糊恢复出清晰图像,其中模糊核的估计是盲去模糊的关键。为了去除纹理和小尺度细节对模糊核估计的干扰,提高模糊核的准确性,提出一个基于结构提取的图像盲去模糊模型。新模型不仅对清晰图像进行图像梯度(边缘)稀疏性约束,还将清晰图像分为结构和纹理两部分,其中结构部分包含图像的光滑部分及其大尺度边缘,然后利用结构部分的梯度代替图像梯度估计来模糊核,最后采取非盲去模糊算法恢复出清晰图像。在多个图像数据集上的实验结果表明,文中方法在定量和视觉效果两方面都显著优于相关方法,在灰度图像数据集上文中方法较具代表性方法所获得去模糊结果的平均PSRN值高1.16dB。Blind image deblurring aims to restore the original sharp image from a blurred observation while the blur kernel is not given.A common method is to estimate blur kernel first,then conducts nonblind deblurring with the aid of the estimated blur kernel to restore clear images.Among them,the blur kernel estimation is the key to blind image deblurring.In order to eliminate the texture and interferences of small-scale details on blur kernel estimation and to improve the accuracy of blur kernel estimation,a novel blind image deblurring model is proposed based on structure extraction.Firstly,the model not only constrains the image gradient(edge)sparsity of the clean image,but also divides the clean image into structure and texture and the structure part contains the smooth of the image and its large-scale edges.Then the structure gradient is used to estimate the blur kernel instead of image gradient.Finally,the clean image is recovered by a nonblind deblurring algorithm.Experimental results on multiple image datasets show that,the quantitative metrics,as well as the visual effects obtained by the proposed method are significantly better than that obtained by the related methods.On a gray image dataset,the average PSRN of the proposed method is 1.16 dB higher than that of the representative methods.

关 键 词：结构提取 盲去模糊 模糊核估计 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]