基于高斯曲率和加权图总变分正则化的遥感图像盲去模糊算法  

Blind remote sensing image deblurring algorithm based on Gaussian curvature and reweighted graph total variation

作　　者：蔡志丹 方明[1] 李喆 许佳路 CAI Zhi-dan;FANG Ming;LI Zhe;XU Jia-lu(School of Artificial Intelligence,Changchun University of Science and Technology,Changchun 130022,China;School of Mathematics and Statistics,Changchun University of Science and Technology,Changchun 130022,China)

机构地区：[1]长春理工大学人工智能学院,长春130022 [2]长春理工大学数学与统计学院,长春130022

出　　处：《吉林大学学报（工学版）》2023年第9期2649-2658,共10页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金项目(12171054);吉林省教育厅科学技术研究项目(JJKH20230791KJ,JJKH20230788KJ)。

摘　　要：针对遥感图像在采集过程中出现的运动模糊现象,本文从图像曲面的几何性质和图像像素的代数性质出发,设计了一种基于高斯曲率和加权图总变分正则化的遥感图像盲去模糊算法。首先,将加权图总变分先验与高斯曲率先验相结合以获得骨架图像,骨架图像保留图像的梯度以及锐利边缘信息,并去除中间潜在图像中的有害结构;然后,利用骨架图像估计模糊核,进而利用非盲去模糊算法获得清晰图像;最后,在8张不同场景下的模糊遥感图像上进行仿真验证。结果表明,相比于其他先进的图像盲去模糊算法,本文提出的去模糊算法复原效果的峰值信噪比平均值分别高于对比算法2.76、1.84、3.11、2.79、3.35、2.76 dB,结构相似性平均值分别高于对比算法0.0792、0.0604、0.0873、0.0801、0.0997、0.0906。本文算法复原的遥感图像具有清晰的边缘轮廓和局部细节,提升了遥感图像的清晰度。To tackle the motion blur in the process of acquiring remote sensing images,an algorithm for blind deblurring of remote sensing images is designed.The algorithm is based on the geometric property of image surfaces and the algebraic property of image pixels,and it utilizes Gaussian curvature and reweighted graph total variation.First,the reweighted graph total variational prior and Gaussian curvature prior were combined to obtain the skeleton image which not only retains the gradient and sharp edge information,but also removes the harmful structural information in the latent clean image.Then,the skeleton image is used to estimate the fuzzy kernel,and then the non-blind deblurring algorithm is used to obtain the clear image.Finally,simulation validation was conducted on 8 fuzzy remote sensing images in different scenarios,and the results showed that,compared with other advanced image deblurring algorithms,the peak signal-tonoise ratio of the recovery effect of the deblurring algorithm proposed is higher than that of the comparison algorithm by 2.76,1.84,3.11,2.79,3.35,2.76 dB,respectively.The structure similarity is higher than that of the comparison algorithm by 0.0792、0.0604、0.0873、0.0801、0.0997、0.0906,respectively.The remote sensing images recovered by our proposed algorithm have clear edge contours and local details while improving the clarity of remote sensing images.

关 键 词：计算数学 遥感图像 高斯曲率 加权图总变分 图像盲去模糊 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]