摘要
针对图像盲去模糊问题,基于变分模型的迭代优化展开形式设计了相应的变分深度学习网络,有效克服了传统变分方法计算效率低和深度学习方法可解释性差的问题。设计网络包含2部分:利用算法展开策略实现基于L_(0)正则化估计模糊核的子网络;基于估计的模糊核及图像恢复正则化模型的非盲去卷积子网络,该子网络充分利用了双通道的编解码网络结构。为确保模糊核估计的准确性和图像内容的一致性,损失函数由均方误差损失和结构相似性损失构成。L_(0)正则化的使用有助于快速准确地完成模糊核估计;图像恢复正则化模型的使用有助于边缘和图像细节的保持。在Levin数据集上的试验结果表明,所提算法在峰值信噪比上较目前先进算法至少提高了2.14 dB。
Aiming at the problem of blind image deblurring,the corresponding variational deep learning network was designed in the form of iterative optimization expansion based on variational model,which effectively overcame the problems of low computational efficiency of traditional variational methods and poor interpretability of deep learning methods.The designed network consisted of two parts:The algorithm unrolling strategy was used to a subnet based on blur kernel estimation of L_(0) regularization terms;A nonblind deconvolution subnet based on an estimated blur kernel and image recovery regularization model,which made full use of the dual-channel coding-decoding network structure.To ensure the accuracy of the blur kernel estimate and the consistency of the image content,the loss function consisted of the loss of mean squared error of the blur kernel and the loss of structural similarity of the image.The use of L_(0) regularization terms facilitated fast and accurate completion of blur kernel estimation;The use of image recovery regularization models helped to maintain edges and image details.Experimental results on Levin dataset showed that the proposed algorithm was at least 2.14 dB higher than the state-of-the-art methods in terms of peak signal-to-noise ratio.
作者
王旭晴
魏伟波
杨光宇
宋金涛
吕婷
潘振宽
WANG Xuqing;WEI Weibo;YANG Guangyu;SONG Jintao;LÜTing;PAN Zhenkuan(College of Computer Science and Technology,Qingdao University,Qingdao 266071,Shandong,China)
出处
《山东大学学报（工学版）》
CSCD
北大核心
2023年第6期35-46,共12页
Journal of Shandong University（Engineering Science）
基金
国家自然科学基金资助项目(61772294)
山东省自然科学基金资助项目(ZR2019LZH002)
山东省高等学校青创科技计划创新团队项目(2021RW018)。
关键词
图像去模糊
算法展开
正则化方法
模糊核估计
深度学习
image deblurring
algorithm unrolling
regularization method
blur kernel estimation
deep learning
