基于全局互补注意力机制的双目图像超分辨率  

Global Complementary Attention Mechanism for Binocular Image Super-resolution Networks

作　　者：鹿存建 杨进华[1] LU Cunjian;YANG Jinhua(School of Opto-Electronics Engineering,Changchun University of Science and Technology,Changchun 130022)

机构地区：[1]长春理工大学光电工程学院,长春130022

出　　处：《长春理工大学学报（自然科学版）》2024年第1期52-59,共8页Journal of Changchun University of Science and Technology(Natural Science Edition)

基　　金：吉林省自然科学基金(YDZJ202301ZYTS396)。

摘　　要：双目图像超分辨率利用双目系统左右图像的互补信息提高图像分辨率,现有方法对跨视图信息的利用率较低。为了解决上述问题,提出一种基于全局互补注意力机制的双目图像超分辨率网络。该网络首先使用多注意力提取模块对视图内的通道和空间特征进行深度提取,然后通过全局互补注意力模块进行视图间的交叉特征提取,最后利用重建模块对特征进行融合。实验结果表明,2倍尺度上未裁剪双目图像的平均峰值信噪比在Middlebury、KITTI2012、KITTI2015和Flickr1024四个基准数据集上分别提高0.76 dB、0.13 dB、0.15 dB、0.38 dB,4倍尺度上分别提高0.18 dB、0.08 dB、0.10 dB、0.10 dB。该网络结构具有较高鲁棒性,还能利用全局信息进行互补,获得更好的主观视觉效果。Binocular image super-resolution utilizes the complementary information of left and right images of binocular system to improve image resolution,and existing methods have low utilization of cross-view information.To address the above problem,a binocular image super-resolution network based on a global complementary attention mechanism is proposed.The network first uses a multi-attention extraction module for deep extraction of channel and spatial features within a view,then performs inter-view cross-feature extraction via a global complementary attention module,and finally fuses the features using a reconstruction module.The experimental results show that the average peak signal-to-noise ratio of the uncropped binocular image on the 2x scale is improved by 0.76 dB,0.13 dB,0.15 dB,0.38 dB on the four benchmark datasets of Middlebury,KITTI 2012,KITTI 2015,and Flickr 1024,r espectively,and on the 4x scale by 0.18 dB,0.08 dB,0.10 dB,and 0.10 dB,respectively.The network structure is highly robust,and it can also utilize the global information for complementary purposes to obtain better subjective visualization.

关 键 词：双目图像 跨视图信息 全局互补注意力 超分辨率 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]