基于极端通道先验和梯度倒谱的图像盲复原  

Blind Image Restoration Based on Extreme Channel Prior and Gradient Cepstrum

作　　者：鱼轮 邢笑笑 YU Lun;XING Xiaoxiao(Artificial Intelligence Research Center,School of Electronic Information and Electrical Engineering,Shangluo University,Shangluo 726000)

机构地区：[1]商洛学院电子信息与电气工程学院人工智能研究中心,商洛726000

出　　处：《舰船电子工程》2024年第10期31-36,共6页Ship Electronic Engineering

基　　金：陕西省教育厅项目(编号:23JK0418);商洛学院科研项目(编号:21SKY010);大学生创新创业训练计划项目(编号:S202311396081)资助。

摘　　要：针对极端通道先验去模糊方法复原得到的图像经常出现振铃伪影等问题,提出一种基于极端通道先验和梯度倒谱的单幅图像盲去模糊方法。首先,对极端通道先验施加l0范数约束并将其引入到最大后验概率框架中构造出模糊核估算模型;然后,对模糊核进行多尺度交替迭代估计,在迭代过程中利用半二次方分裂法有效解决模型的非凸问题。为了抑制每个尺度的过度迭代,利用核相似度来评估迭代过程中的模糊核细微变化,从而使得最终迭代得到的模糊核更加精确。最后,通过非盲解卷积实现图像的去模糊。实验表明,所提方法在合成数据集与真实数据集上取得了良好效果,能够抑制伪影和恢复出更多的图像细节。Aiming at the problems of ringing artifacts in the images recovered by the extreme channel prior deblurring method,a single image blind deblurring method based on the extreme channel prior and gradient cepstrum is proposed.Firstly,the l0 norm constraint is imposed on the extreme channel prior and introduced into the maximum a posteriori probability framework to construct the fuzzy kernel estimation model.Then,the fuzzy kernel is estimated by multi-scale alternating iterative estimation,and the semi-quadratic splitting method is used to effectively solve the non-convex problem of the model in the iterative process.In order to inhibit the excessive iteration of each scale,the kernel similarity is used to evaluate the subtle changes of the fuzzy kernel in the iterative process,so that the final iterative fuzzy kernel is more accurate.Finally,the image deblurring is realized by non-blind deconvolution.Experimental results show that the proposed method achieves good results on both synthetic and real datasets,which can suppress artifacts and recover more image details.

关 键 词：极端通道先验 梯度倒谱 模糊核 多尺度 半二次方分裂法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]