基于图像超分辨率的多尺度特征交互传播研究  

Research on Multi-scale Feature Propagation and Communication Based on Image Super Resolution

作　　者：孙有晓 冯锐成 关伟鹏 乔宇[2] 董超 井焜 刘辰飞 许野平 陈英鹏 周卫东 SUN Youxiao;FENG Ruicheng;GUAN Weipeng;QIAO Yu;DONG Chao;JING Kun;LIU Chenfei;XU Yeping;CHEN Yingpeng;ZHOU Weidong(School of Microelectronies,Shandong University,Jinan 250101,China;Shenzhen Institute of Advanced Technology,Chinese Academy of Sciences,Shenzhen 518055,China;Shandong Institute of Advanced Technology Co.,Ltd.,Chinese Academy of Sciences,Jinan 250101,China;Synthesis Electronics Technology Co.,Ltd.,Jinan 250101,China)

机构地区：[1]山东大学微电子学院,济南250101 [2]中国科学院深圳先进技术研究院,深圳518055 [3]山东中科先进技术研究院有限公司,济南250101 [4]神思电子技术股份有限公司,济南250101

出　　处：《集成技术》2022年第2期41-54,共14页Journal of Integration Technology

基　　金：山东省发展和改革委员会新旧动能转换重大工程重大课题攻关项目;山东省自然科学基金联合基金重点项目(ZR2020LZH009)~~。

摘　　要：该文构建了一种可对不同形式的多尺度结构进行归纳的统一框架,并基于该框架系统地探究了多尺度卷积的两个因素——特征传播和跨尺度交互,提出了简单而有效的多尺度卷积单元——多尺度-跨尺度-权重共享的卷积(MS;-Conv)网络。实验结果表明,与基于标准卷积的网络相比,基于MS;-Conv的网络可使用较少的参数和较低的计算成本实现更好的图像重建性能。除了定量分析,该文也对重建图像进行可视化分析,证明了MS;-Conv网络能更好地恢复高频细节。This paper presents a unified framework for various multi-scale structures.With this framework,two factors of multi-scale convolution,i.e.feature propagation and cross-scale communication are explored.A generic and efficient multi-scale convolution unit named Multi-Scale cross-Scale Share-weights Convolution(MS;-Conv)is proposed.Experimental results showed that,the proposed MS;-Conv can achieve better super resolution performance than conventional convolution methods with less parameters and computational cost.By observation of the visual quality,results also showed that the MS;-Conv outperform in the reconstruction of high-frequency image details.

关 键 词：深度学习 多尺度卷积 超分辨率 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]