摘要
针对单幅图像超分辨率重建任务中单域特征的重建能力受限以及深度卷积神经网络参数众多、计算量大导致的难以部署到移动端的问题,提出了一种基于多域信息增强的轻量级图像超分辨率网络。从空域、频域和转换域3个维度入手,设计了空域多路大核特征提取技术、局部信息增强注意力、频域分频特征增强技术以及转换域基于高频特征模拟技术。通过不同特征域的信息处理,针对全局与局部的低频和高频特征进行精准优化,从而提升模型在细节恢复与图像重建中的表现。与现有先进算法在公认基准数据集上进行充分的实验对比和分析,结果表明所提网络模型能够实现优异的重建效果,且在性能与效率之间也实现了出色的平衡。
Aiming to solve the problems that the reconstruction capability of single-domain features was limited and deep convolutional neural networks used in existing single-image super-resolution reconstruction tasks were difficult to deploy on mobile terminals due to the large number of parameters and high computational requirements,a lightweight image super-resolution network based on multi-domain information enhancement was proposed.Initiating from three dimensions,a set of innovative techniques had been developed,including multi-path large kernel feature extraction in the spatial domain,local information enhancement attention,frequency-domain feature enhancement through frequency splitting,and transformation-domain prior-guided high-frequency feature simulation.By processing information across different feature domains,both global and local low-frequency and high-frequency features were optimized,significantly improving the model’s performance in detail recovery and image reconstruction.Extensive experimental comparisons and analyses with the existing advanced algorithms on the recognized benchmark datasets demonstrate that the proposed network model can achieve remarkable reconstruction results while enjoying a high trade-off between performance and efficiency.
作者
寇旗旗
刘规
江鹤
陈亮亮
程德强
KOU Qiqi;LIU Gui;JIANG He;CHEN Liangliang;CHENG Deqiang(School of Computer Science and Technology,China University of Mining and Technology,Xuzhou 221116,China;School of Information and Control Engineering,China University of Mining and Technology,Xuzhou 221116,China)
出处
《通信学报》
北大核心
2025年第4期144-159,共16页
Journal on Communications
基金
国家自然科学基金资助项目(No.52204177,No.52304182)
济宁市重点研发基金资助项目(No.2021KJHZ013,No.2023KJHZ007)。
关键词
计算机视觉
超分辨率
多域信息增强
注意力
轻量级
computer vision
super resolution
multi-domain information enhancement
attention
lightweight
作者简介
寇旗旗(1988-),男,河南襄城人,博士,中国矿业大学副教授、硕士生导师,主要研究方向为图像处理/AI视频分析、智能检测与模式识别、图像增强与复原;刘规(1999-),男,重庆酉阳人,中国矿业大学硕士生,主要研究方向为图像处理、图像超分辨率重建;江鹤(1990-),男,江苏徐州人,博士,中国矿业大学讲师,主要研究方向为图像修复、图像识别和三维重建;陈亮亮(1993-),男,江苏徐州人,博士,中国矿业大学讲师,主要研究方向为图像处理、图像增强和超分辨;通信作者:程德强(1979-),男,河南洛阳人,博士,中国矿业大学教授、博士生导师,主要研究方向为图像智能检测与模式识别、图像处理与视频编码,chengdq@cumt.edu.cn。
