基于类金字塔图残差网络的图像超分辨率重建  

Pyramid-like graph residual network for image super-resolution reconstruction

作　　者：赵小强[1,2,3] 王泽 宋昭漾 蒋红梅[1,2,3] ZHAO Xiao-qiang;WANG Ze;SONG Zhao-yang;JIANG Hong-mei(College of Electrical Engineering and Information Engineering,Lanzhou University of Technology,Lanzhou 730050,China;Key Laboratory of Gansu Advanced Control for Industrial Processes,Lanzhou 730050,China;National Experimental Teaching Center of Electrical and Control Engineering,Lanzhou University of Technology,Lanzhou 730050,China)

机构地区：[1]兰州理工大学电气工程与信息工程学院,兰州730050 [2]甘肃省工业过程先进控制重点实验室,兰州730050 [3]兰州理工大学国家级电气与控制工程实验教学中心,兰州730050

出　　处：《控制与决策》2024年第3期786-794,共9页Control and Decision

基　　金：国家自然科学基金项目(62263021);国家重点研发计划项目(2020YFB1713600);甘肃省科技计划项目(21YF5GA072,21JR7RA206)。

摘　　要：针对基于深度学习的图像超分辨率重建算法大多侧重于从大量外部训练数据中学习,而忽视图像本身的内部知识以及过于关注局部特征的问题,提出一种基于类金字塔图残差网络的图像超分辨率重建算法.首先,该算法构建的残差图卷积结构利用一种预生成图结构的方式将提取的特征图转换为预生成图结构的顶点来构成图结构数据,从而通过图卷积来学习特征自身内部的拓扑结构,同时使用残差学习适度地加深图卷积网络以提高重建性能;其次,该算法构建的类金字塔多空洞卷积结构,通过充分利用不同大小的感受野,避免了不能完全覆盖所有像素点的缺陷,更好地融合不同尺度的特征信息;最后,经过大量实验验证,所提出的算法显著优于主流超分辨率方法,有着更好的客观和主观度量结果.To address the problems that the most of deep learning-based super-resolution reconstruction algorithms focus on learning from a large amount of external training data while ignoring the internal knowledge of an image itself and focusing too much on local features,a pyramid-like graph residual network is proposed for image super-resolution reconstruction.Firstly,the algorithm builds a residual graph convolution structure,which converts the extracted feature maps into vertices of pre-generated graph structures to constitute graph structure data by using a kind of pre-generated graph structure,so as to learn the internal topology of the features themselves by graph convolution,and the residuals are used to learn a moderate deepening graph convolution network to improve the reconstruction performance.Then,the algorithm builds a pyramid-like multi-dilated convolution structure,which avoids the defect of not completely covering all pixel points by making full use of different sizes of perceptual fields and better fuses feature information at different scales.Finally,experimental results show that the proposed algorithm significantly outperforms the mainstream super-resolution algorithms with better objective and subjective metric results.

关 键 词：图像复原 超分辨率重建 图卷积网络 空洞卷积 残差学习 金字塔结构 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]