基于层次注意力的图像超分辨率重建  

Image Super-resolution Reconstruction Based on Hierarchical Attention

作　　者：臧永盛 周冬明[1] 王长城 聂仁灿[1] ZANG Yongsheng;ZHOU Dongming;WANG Changcheng;NIE Rencan(School of Information Science&Engineering,Yunnan University,Kunming 650504,China)

机构地区：[1]云南大学信息学院,云南昆明650504

出　　处：《无线电工程》2021年第11期1245-1253,共9页Radio Engineering

基　　金：国家自然科学基金资助项目(62066047,61966037,61365001,61463052)。

摘　　要：卷积神经网络的深度和提供的层次特征对于超分辨率图像的重建至关重要,盲目增加网络的深度会使网络结构过于复杂且图像高频信息容易丢失。对低分辨率图像中提供的层次特征同等地对待而不加区分,会阻碍卷积神经网络的表达能力。针对该问题,提出了一种基于多尺度残差和层次注意力的重建网络。网络由不同层次的跳过连接、多尺度残差块和层次注意力模块组成,能实现有效的图像重建;利用不同层次的跳过连接绕过图像中大量低频信息,使主网络专注于学习图像高频信息。设计了多尺度残差模块和层次注意力模块以自适应地提取不同比例的图像特征和有针对性地融合不同层次的特征。实验结果表明,与其他模型相比,所提出的模型更能有效地利用原始图像信息,恢复出细节更清晰的超分辨率图像。提出的模型在客观评价指标上表现优异且模型参数数目少,易于训练。The depth of the convolutional neural network and the hierarchical features provided are crucial for the reconstruction of super-resolution images.However,blindly increasing the depth of the network can make the network structure too complex and the high-frequency information of the image easily lost.In addition,treating the hierarchical features provided in low-resolution images equally without differentiation can hinder the expressive ability of convolutional neural networks.Therefore,a reconstruction network based on multi-scale residual and hierarchical attention is proposed.The network consists of different levels of skip connections,multi-scale residual blocks and hierarchical feature fusion modules,which can achieve effective image reconstruction.The network uses different levels of skip connections to bypass a large amount of low-frequency information in the image,allowing the main network to focus on learning high-frequency information of the image.In addition,the network is designed with a residual module and a hierarchical attention module to adaptively extract image features at different scales and fuse features at different levels in a targeted manner.The experimental results show that the proposed model is more effective in using the original image information to recover super-resolution images with clearer details than other models.In addition,the proposed model performs well in objective evaluation metrics and is easy to train with a small number of model parameters.

关 键 词：超分辨率 卷积神经网络 多尺度 残差网络 层次特征 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]