基于深度先验的盲图像去模糊算法  

Deep Prior-Based Blind Image Deblurring

作　　者：白勇强 禹晶 李一秾 肖创柏 BAI Yong-qiang;YU Jing;LI Yi-nong;XIAO Chuang-bai(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China)

机构地区：[1]北京工业大学信息学部,北京100124

出　　处：《电子学报》2023年第4期1050-1067,共18页Acta Electronica Sinica

基　　金：北京市自然科学基金(No.4212014);北京市教育委员会科技计划(No.KM201910005029)。

摘　　要：盲图像去模糊旨在模糊核未知的情况下从模糊图像恢复清晰图像,这是一个欠定逆问题,需要引入图像先验信息限定解空间.受到SelfDeblur的启发,本文提出了一种基于深度先验的盲图像去模糊算法,结合深度网络与正则化模型对清晰图像与模糊核联合建模,交替迭代估计清晰图像与模糊核.在图像估计子问题中,模糊核参与RGB三通道损失函数的约束下,利用隐含图像平滑性约束的深度卷积神经网络DIP-Net生成清晰图像;在模糊核估计子问题中,直接求取模糊核正则化约束模型的全局极小解,不同于SelfDeblur的全连接网络使用梯度下降法更新模糊核.本文算法结合深度网络实现正则化方法,与监督学习相比,无需成对的模糊/清晰图像数据集训练网络;与传统模型方法相比,无需通过多级金字塔的方式由粗到细地估计模糊核.在模拟与真实模糊图像上的实验结果表明;本文算法能够快速、准确地估计出清晰图像和模糊核,并能够有效抑制图像复原过程中存在的噪声放大问题.Blind image deblurring is the process of removing blurring artifacts from the observation when the blur kernel is unknown,which is a seriously ill-posed problem.It is indispensable to impose prior information constraints on the feasible set.Inspired by the SelfDeblur,in this paper we propose a deep prior-based blind image deblurring method,which uses the deep network and the regularized optimization model to jointly optimize and alternately update the latent image and the blur kernel.Conditioned by the loss of the sum of RGB three-channel errors with the presence of blur kernel,the latent im-age is estimated using the deep convolutional neural network DIP-Net implicitly involving the smoothness regularizer of the image.The blur kernel estimation subproblem admits the global optimal solution,which is different from the SelfDeblur that applies the fully connected network and takes the gradient descent step to update the blur kernel.Our method uses the struc-ture of the deep network to regularize the latent image.Unlike the supervised image deblurring method,it requires no ground truth of the latent image or the blur kernel.Unlike the traditional model,it requires no progressive transmission from coarse to fine through the multi-level pyramid.Experimental results on simulated and real blur images show that the proposed method achieves a fast and accurate estimation of both the blur kernel and the latent image with efficient noise suppression.

关 键 词：盲图像去模糊 深度先验 卷积神经网络 正则化 盲解卷积 图像复原 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]