基于残差注意力机制的图像超分辨率重建  

Image Super-Resolution Reconstruction Based on Residual Attention Mechanism

作　　者：李岚 张云 何方 尹喆 LI Lan;ZHANG Yun;HE Fang;YIN Zhe(School of Digital Media,Lanzhou University of Arts and Science,Lanzhou 730000,China;Network Center,Lanzhou Radio and Television Station,Lanzhou 730031,China)

机构地区：[1]兰州文理学院数字媒体学院,甘肃兰州730000 [2]兰州广播电视台网络中心,甘肃兰州730031

出　　处：《兰州文理学院学报（自然科学版）》2024年第5期48-53,共6页Journal of Lanzhou University of Arts and Science(Natural Sciences)

基　　金：甘肃省自然科学基金项目(23JRRA746,24JRRA162);兰州市科技计划项目(2023-3-120)。

摘　　要：针对传统方法中纹理、边缘、遮挡等区域难以进行图像超分辨率重建的问题,提出一种基于注意力机制的残差网络超分辨率重建方法.首先对输入图像应用一层卷积网络和一层位移网络提取浅层特征;然后引入若干个残差坐标注意力模块自适应校正通道权重,与浅层特征相加提取深层特征;最后,结合亚像素卷积重建出高分辨率图像.在标准数据集Set5和Set14上的实验结果表明,该方法重建图像的客观评价指标和视觉效果均优于SRCNN、VDSR和Bicubic方法,在一定程度上改善了重建图像的质量,并提升了算法的运行速度.A residual network super-resolution reconstruction method based on coordinate attention mechanism is proposed to address the problem of difficulty in image reconstruction of texture,edge,occlusion and other areas in traditional methods.Firstly,a 1-layer convolutional network and 1-layer displacement are applied to extract shallow features from the input image.Then,several residual coordinate attention modules are introduced to adaptively correct channel weights and extract deep features by adding them to shallow features.Finally,high-resolution images are reconstructed by combining sub-pixel convolution.The experimental results on the standard datasets Set5 and Set14 show that the objective evaluation metrics and visual effects of the reconstructed images using this method are superior to those of SRCNN,VDSR,and LapSRN methods,which improves the quality of the reconstructed images to a certain extent and enhances the running speed of the algorithm.

关 键 词：卷积神经网络 残差注意力机制 特征提取 超分辨率重建 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]