残差注意力与多特征融合的图像去模糊  

Image Deblurring Based on Residual Attention and Multi-feature Fusion

作　　者：赵倩 周冬明[1] 杨浩 王长城 ZHAO Qian;ZHOU Dongming;YANG Hao;WANG Changchen(School of Information Science&Engineering,Yunnan University,Kunming 650504,China)

机构地区：[1]云南大学信息学院,昆明650504

出　　处：《计算机科学》2023年第1期147-155,共9页Computer Science

基　　金：国家自然科学基金(61966037,62066047)。

摘　　要：动态场景下的非均匀盲去模糊是一个极具挑战性的计算机视觉问题。虽然基于深度学习的去模糊算法已经取得很大进展,但仍存在去模糊不彻底和细节丢失等问题。针对这些问题,提出了一种基于残差注意力和多特征融合的去模糊网络。与现有的单分支网络结构不同,所提网络由两个独立的特征提取子网组成。主干网络采用基于U-Net结构的编码器-解码器网络来获取不同层级的图像特征,并使用残差注意力模块对特征进行筛选,从而自适应地学习图像的轮廓特征和空间结构特征。另外,为了补偿主干网络中下采样操作和上采样操作造成的信息损失,进一步利用具有大感受野的深层次加权残差密集子网来提取特征图的细节信息。最后,使用多特征融合模块逐步融合原分辨率模糊图像以及主干网络和加权残差密集子网生成的特征信息,使得网络能够以整体的方式自适应地学习更有效的特征来复原模糊图像。为了评估网络的去模糊效果,在基准数据集GoPro数据集和HIDE数据集上进行了测试,结果表明所提方法能够有效复原模糊图像。与现有方法相比,提出的去模糊算法在视觉效果上和客观评价指标上均取得了很好的去模糊效果。Non-uniform blind deblurring in dynamic scenes is a challenging computer vision problem.Although deblurring algorithms based on deep learning have made great progress,there are still problems such as incomplete deblurring and loss of details.To solve these problems,a deblurring network based on residual attention and multi-feature fusion is proposed.Unlike the existing single-branch network structure,the proposed network consists of two independent feature extraction subnets.The backbone network uses an encoder-decoder network based on U-Net to obtain image features at different scales,and uses the residual attention module to filter the features,so as to adaptively learn the contour features and spatial structure features of the image.In addition,in order to compensate for the information loss caused by the down-sampling operation and up-sampling operation in the backbone network,a deep weighted residual dense subnet with a large receptive field is further used to extract rich detailed information of the feature map.Finally,the multi-feature fusion module is used to gradually fuse the original resolution blurred image and the feature information generated by the backbone network and the weighted residual dense subnet,so that the network can adaptively learn more effective features in an overall manner to restore the blurred image.In order to evaluate the deblurring performance of the network,tests are conducted on the benchmark data sets GoPro and HIDE,and the results show that the blurred image can be effectively restored.Compared with the existing methods,the proposed deblurring algorithm has achieved excellent deblurring performances in terms of visual effects and objective evaluation indicators.

关 键 词：图像去模糊 注意力机制 编码-解码结构 密集残差网络 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]