Efficient single-image super-resolution:deeply-supervised symmetric distillation network

摘要：

过去几十年,卷积神经网络(Convolutional Neural Networks,CNNs)在单图像超分辨率(Single Image Super-Resolution,SISR)方面取得了明显的进展.现在大部分基于CNNs的方法都致力于构造新的架构去提升重建性能,这通常依赖大量计算和存储成本,难以应用于移动设备.提出了一种新颖的基于深监督对称蒸馏网络的高效单图像超分辨率重建方法(Deeply-Supervised Symmetry Distillation Network,DSSD),通过构造高频特征递归模块(High-frequency Feature Recursive Module,HFRM)和对称退化模块(Symmetry Degradation Module,SDM)缓解教师网络中提取高分辨率(High-Resolution,HR)高频信息不够准确这一问题.为了约束教师网络中提取的高频特征,采用深监督方法使教师网络蒸馏的知识与学生网络互补.在DIV2K数据集上的实验表明,DSSD有效增强了单图像超分辨率(SISR)的性能,HFRM和SDM的引入能够有效帮助DSSD提取更多图像高频细节.

Convolutional neural networks(CNNs)have made significant advances in single image super resolution(SISR)over the past few decades.Most CNNs-based approaches nowadays are devoted to constructing new architectures to improve reconstruction performance,which usually rely on large computation and storage costs and are difficult to apply to mobile devices.In this paper,we propose a novel efficient single-image super-resolution reconstruction method(DSSD)based on deeply supervised symmetric distillation networks,the problem that the high-frequency information of High-Resolution extracted in the teacher network is not accurate enough is alleviated by constructing a high-frequency feature recursion module(HFRM)and a symmetric degradation module(SDM).To constrain the high frequency features extracted from the teacher network,a deep supervision approach was used to make the knowledge distilled from the teacher network complementary to the student network.Experiments on the DIV2K dataset show that DSSD effectively enhances the performance of SISR,and the introduction of HFRM and SDM can effectively help DSSD to extract more high-frequency details of images.

作者：

毛盼娣 徐道连

Mao Pandi;Xu Daolian(School of Electrical Engineering and Intelligent Manufacturing,Chongqing Metropolitan College of Science and Technology,Chongqing 402167,China;School of Optoelectronic Engineering,Chongqing University,Chongqing 400044,China)

机构地区：

重庆城市科技学院电气工程与智能制造学院

出处：

《河南师范大学学报:自然科学版》 CAS 北大核心  2023年第6期57-64,I0002,共9页

Journal of Henan Normal University(Natural Science Edition)

基金：

重庆市教委科学技术研究项目(KJQN202002501) 重庆市高等教育教学改革研究项目(213473).

关键词：

深监督对称蒸馏网络 超分辨率 教师网络 高频特征递归模块 对称退化模块 特权信息

deeply-supervised symmetric distillation networks super-resolution teacher network high-frequency feature recursion module symmetric degeneration module privileged information

分类号：

TP391.41 [自动化与计算机技术—计算机应用技术] 

高效单图像超分辨率重建：深监督对称蒸馏网络.pdf