基于双注意力对偶学习的图像超分辨率重建算法  

Image super-resolution reconstruction algorithm based on dual attention dual learning

作　　者：刘恒 梁媛 潘斌[2] 刘琳珂 侯健 LIU Heng;LIANG Yuan;PAN Bin;LIU Linke;HOU Jian(School of Information and Control Engineering,Liaoning Petrochemical University,Fushun 113001,China;School of Sciences,Liaoning Petrochemical University,Fushun 113001,China)

机构地区：[1]辽宁石油化工大学信息与控制工程学院,辽宁抚顺113001 [2]辽宁石油化工大学理学院,辽宁抚顺113001

出　　处：《辽宁科技大学学报》2022年第2期119-126,共8页Journal of University of Science and Technology Liaoning

基　　金：国家自然科学基金(61602228、61572290);辽宁省“兴辽英才计划”青年拔尖人才项目(XLYC1807266);辽宁省教育厅项目(L2020018)。

摘　　要：针对单幅图像低分辨率到高分辨率映射的不适定性,以及特征图通道域和空间域信息利用率低的问题,本文引入基于双注意力机制的对偶学习算法,用于单幅图像超分辨率的重建。算法先对输入图像进行特征提取,较大程度保留特征信息;之后采用双注意力机制计算图像的通道域和空间域的显著性,以提取到更准确有效的深层特征;最后利用对偶学习构建闭环反馈网络,通过对偶关系约束映射空间,以获取最优重建函数。在基准数据集Set5、Set14、BSDS100、Urban100上进行放大2倍和4倍的重建测试实验表明,与其他超分辨率算法相比,本文算法的峰值信噪比和结构相似度都高于其他算法,其视觉效果也比其他算法的图像更清晰。Aiming at the ill-posedness mapping from low-resolution to high-resolution of single image,and the low information utilization in the channel domain and spatial domain of the feature map,an super-resolution reconstruction algorithm for single image based on dual attention dual learning is introduced. Firstly,the features of the input image are extracted,and the feature information is retained to a large extent. Secondly,the saliency of the channel domain and spatial domain of the image is calculated based on the dual attention mechanism to extract more accurate and effective deep features. Finally,a closed-loop feedback network is constructed by dual learning,and the mapping space is constrained by the dual relationship to obtain the optimal reconstruction function. Reconstruction test experiments with 2x and 4x magnification on benchmark datasets Set5,Set14,BSDS100,Urban100 showed that compared with other super-resolution algorithms,the peak signal-to-noise ratio and structural similarity of the proposed algorithm were higher than other algorithms. From the visual effect,the image was clearer than other algorithms.

关 键 词：超分辨率重建 双注意力机制 对偶学习 闭环反馈网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]