基于组−信息蒸馏残差网络的轻量级图像超分辨率重建  

G-IDRN:A Group-information Distillation Residual Network for Lightweight Image Super-resolution

作　　者：王云涛 赵蔺 刘李漫[1] 陶文兵[2] WANG Yun-Tao;ZHAO Lin;LIU Li-Man;TAO Wen-Bing(School of Biomedical Engineering,South-central Minzu University,Wuhan 430074;School of Artificial Intelligence and Automation,Huazhong University of Science and Technology,Wuhan 430074)

机构地区：[1]中南民族大学生物医学工程学院,武汉430074 [2]华中科技大学人工智能与自动化学院,武汉430074

出　　处：《自动化学报》2024年第10期2063-2078,共16页Acta Automatica Sinica

基　　金：国家自然科学基金(61976227,62176096);湖北省自然科学基金(2019CFB622)资助。

摘　　要：目前,基于深度学习的超分辨算法已经取得了很好性能,但这些方法通常具有较大内存消耗和较高计算复杂度,很难应用到低算力或便携式设备上.为了解决这个问题,设计一种轻量级的组−信息蒸馏残差网络(Group-information distillation residual network,G-IDRN)用于快速且精确的单图像超分辨率任务.具体地,提出一个更加有效的组−信息蒸馏模块(Group-information distillation block,G-IDB)作为网络特征提取基本块.同时,引入密集快捷连接,对多个基本块进行组合,构建组−信息蒸馏残差组(Group-information distillation residual group,G-IDRG),捕获多层级信息和有效重利用特征.另外,还提出一个轻量的非对称残差Non-local模块,对长距离依赖关系进行建模,进一步提升超分性能.最后,设计一个高频损失函数,去解决像素损失带来图像细节平滑的问题.大量实验结果表明,该算法相较于其他先进方法,可以在图像超分辨率性能和模型复杂度之间取得更好平衡,其在公开测试数据集B100上,4倍超分速率达到56 FPS,比残差注意力网络快15倍.Recently,most super-resolution algorithms based on deep learning have achieved satisfactory results.However,these methods generally consume large memory and have high computational complexity,and are difficult to apply to low computing power or portable devices.To address this problem,this paper introduces a lightweight group-information distillation residual network(G-IDRN)for fast and accurate single image super-resolution.Specially,we propose a more effective group-information distillation block(G-IDB)as the basic block for feature extraction.Simultaneously,we introduce dense shortcut to combine them to construct a group-information distillation residual group(G-IDRG),which is used to capture multi-level information and effectively reuse the learned features.Moreover,a lightweight asymmetric residual Non-local block is proposed to model the long-range dependencies and further improve the performance of super-resolution.Finally,a high-frequency loss function is designed to alleviate the problem of smoothing image details caused by pixel-wise loss.Extensive experiments show the proposed algorithm achieves a better trade-off between image super-resolution performance and model complexity against other state-of-the-art super-resolution methods and gets 56 FPS on the public test dataset B100 with a scale factor of 4 times,which is 15 times faster than the residual channel attention network.

关 键 词：残差网络 超分辨率 特征蒸馏 高频损失 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]