基于显著性先验的自适应空可变图像去模糊方法  

Adaptive space-variant single image deblurring method based on saliency map

作　　者：时愈[1,2,3] 严嘉倩 黄志高 华夏[1,2,3] SHI Yu;YANJiaqian;HUANG Zhigao;HUA Xia(School of Electrical and Information Engineering,Wuhan Institute of Technology,Wuhan 430205,China;Hubei Key Laboratory of Optical Information and Pattern Recognition,Wuhan Institute of Technology,Wuhan 430205,China;Laboratory of Hubei Province Video Image and HD Projection Engineering Technology Research Center,Wuhan Institute of Technology,Wuhan 430205,China)

机构地区：[1]武汉工程大学电气信息学院,湖北武汉430205 [2]武汉工程大学光学信息与模式识别湖北省重点实验室,湖北武汉430205 [3]武汉工程大学湖北省视频图像与高清投影研究中心,湖北武汉430205

出　　处：《华中科技大学学报（自然科学版）》2021年第9期30-35,46,共7页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61701353,61801337)。

摘　　要：针对图像分块不够合理导致的模糊核估计不精确或计算量较大等问题,提出了一种从单幅空间可变模糊图像中自适应估计清晰图像的方法.首先,采用四叉树分解将空可变模糊图像进行自适应划分,按照模糊核的相似程度确保划分的不同图像块具有不同的模糊核.然后,对于每一个图像块,施加了一种显著性值先验来恢复潜在清晰图像,提取人眼感兴趣的重要部分能够更精确地估计模糊核,并且可以保留更多的图像细节.最后,采用加权窗函数对尺寸大小不同的相邻图像块进行拼接,得到去模糊后的复原图像.实验结果表明:在对空不变和空可变的模糊图像进行去模糊后,与目前的经典方法相比,提出的方法可以得到较好的图像恢复结果.The irrationality of image dividing will lead to the problems of the inaccurate kernel estimation and excessive computation.In order to solve this problem,an efficient method for adaptively recovering a single spatially variant blurred image was proposed.Based on the blur kernels’similarity to ensure that the obtained patches have different kernels,the quadtree decomposition was introduced to adaptively partition the image.Then a saliency map was added to recover the latent image of each patch,which is capable extracting the essential parts that human eyes are interested to estimate kernel more accurately and preserve more image details.Finally,a weighted window function was combined to joint adjacent patches with different sizes to get the deblurring result.The experimental results on both space-invariant and space-variant images show that our proposed method can achieve the comparable or better restoration compared with the existing classic methods.

关 键 词：图像分块 自适应去模糊 四叉树分解 显著性图 加权窗函数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]