基于局部分离与多尺度融合的图像超分辨率重建  

Image Super-Resolution Reconstruction Based on Partial Separation and Multiscale Fusion

作　　者：杨郅树 梁佳楠 曹永军 钟震宇 何永伦 YANG Zhishu;LIANG Jianan;CAO Yongjun;ZHONG Zhenyu;HE Yonglun(School of Faculty of Intelligent Manufacturing,Wuyi University,Jiangmen 529020,Guangdong,China;Guangdong Key Laboratory of Modern Control Technology,Institute of Intelligent Manufacturing,Guangdong Academy of Sciences,Guangzhou 510070,Guangdong,China;School of Mechanical&Automotive Engineering,South China University of Technology,Guangzhou 511442,Guangdong,China)

机构地区：[1]五邑大学智能制造学部,广东江门529020 [2]广东省科学院智能制造研究所广东省现代控制技术重点实验室,广东广州510070 [3]华南理工大学机械与汽车工程学院,广东广州511442

出　　处：《计算机工程》2024年第7期314-323,共10页Computer Engineering

基　　金：佛山市重点领域科技攻关项目(2020001006827);广州市科技计划项目(202206010052)。

摘　　要：目前基于深度学习的超分辨率重建网络存在卷积运算冗余、图像重建信息不完整、模型参数庞大等问题,限制了其在边缘设备上的适用性。针对上述问题,提出一种轻量级的局部分离与多尺度融合图像超分辨率重建网络,该网络利用局部卷积对图像进行特征提取,通过分离部分图像通道,在减少网络冗余计算的同时保持图像重建的质量。设计一种多尺度特征融合模块,在空间维度学习长依赖特征,并采用一个通道注意力增强组在空间维度捕获空间特征,减少图像重建信息的丢失,有效恢复图像的细节纹理。由于多尺度特征融合模块更多地是从全局角度进行特征提取融合,因此构建一种高效反残差模块补充网络的局部上下文信息提取能力。在Set5、Set14、B100、Urban100、Manga109这5个基准数据集上的实验结果表明,当尺度因子为2、3、4倍时,该网络的参数量分别为373000、382000、394000,FLOPs分别为84.0×10^(9)、38.1×10^(9)、22.1×10^(9)。与VDSR、IMDN、RFDN、RLFN等网络相比,该网络在较少网络参数的情况下,能够保证图像重建效果。Currently,deep-learning-based super-resolution reconstruction networks suffer from issues such as convolution operation redundancy,incomplete image reconstruction information,and large model parameters that limit their applicability to edge devices.To address these issues,this study proposes a lightweight image superresolution reconstruction network based on partial separation and multiscale fusion.This network utilizes partial convolutions for feature extraction and separates partial image channels to reduce redundant computations while maintaining the quality of the image reconstruction.At the same time,a multiscale feature fusion module is designed to learn long-range dependency features and capture spatial features in the spatial dimension using a channel attention enhancement group.This reduces the loss of image reconstruction information and effectively restores the details and textures of the image.Finally,because the multiscale feature fusion block focuses on global feature extraction and fusion,an efficient inverted residual block is constructed to supplement the ability to extract local contextual information.The network is tested on five benchmark datasets:Set 5,Set 14,B 100,Urban 100,and Manga 109,with scale factors of 2,3,and 4 times.The parameters of the network are 373000,3820000,and 394000,and the FLOPs are 84.0×10^(9),38.1×10^(9),and 22.1×10^(9),respectively.Quantitative and qualitative experimental results show that compared with networks such as VDSR,IMDN,RFDN,and RLFN,the proposed network ensures image reconstruction quality with fewer network parameters.

关 键 词：超分辨率重建 轻量级网络 局部卷积 多尺度融合 长依赖关系 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]