自适应深度残差椒盐噪声滤除算法  

Adaptive Salt-and-Pepper Denoising Based on Deep Residual Network

作　　者：邓森 徐进轩 梁鹿鸣 杨珉 谢浩然 王富利 汪俊 魏明强[1,2,3] 郭延文 Deng Sen;Xu Jinxuan;Liang Luming;Yang Min;Xie Haoran;Wang Fuli;Wang Jun;Wei Mingqiang;Guo Yanwen(College of Computer Science&Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 210016;MIIT Key Laboratory of Pattern Analysis and Machine Intelligence,Nanjing 210016;State Key Laboratory for Novel Software Technology,Nanjing University,Nanjing 210023;Microsoft Research,Seattle 90876;College of Software,Jilin University,Changchun 130012;Department of Computing and Decision Sciences,Lingnan University,Hong Kong 999077;College of Science and Technology,The Open University of Hong Kong,Hong Kong 999077)

机构地区：[1]南京航空航天大学计算机科学与技术学院,南京210016 [2]模式分析与机器智能工业和信息化部重点实验室,南京210016 [3]南京大学计算机软件新技术国家重点实验室,南京210023 [4]Microsoft Research,Seattle 908765 [5].吉林大学软件学院,长春130012 [6]香港岭南大学电脑及决策科学学系,中国香港999077 [7]香港公开大学科技学院,中国香港999077

出　　处：《计算机辅助设计与图形学学报》2020年第8期1248-1257,共10页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61502137);中央高校基本科研业务费(NJ2019010);南京大学计算机软件新技术国家重点实验室开放课题(KFKT2018B20);香港岭南大学香港商学研究所2019-20种子研究基金(190-009);香港岭南大学种子研究基金(102367);香港岭南大学陈斌博士数据科学机构项目(LEO Dr David P.Chan Institute of DataScience);南京航空航天大学研究生创新基地(实验室)开放基金(Kfjj20191601)。

摘　　要：为了在去除图像中椒盐噪声的同时最大程度地避免产生色彩失真与边缘模糊等瑕疵,提出基于深度残差网络的椒盐噪声自适应滤除算法.将图像去噪分解为2步.首先,为了让网络模型能够处理不同尺度密度的椒盐噪声,提高网络模型的鲁棒性,先对图像进行自适应预处理以去除高频信息;其次,构建深度残差网络模型,训练出能将预处理后的图像映射到干净图像的函数.大量实验结果表明,文中算法不仅在保留图像边缘细节和去除高密度椒盐噪声方面均优于传统和基于机器学习的椒盐噪声去除技术,可有效地避免出现色彩失真和条纹等瑕疵.同时,其在BSD300数据集上去噪效果优于其他算法.To remove salt-and-pepper noise with minimal degradation(e.g.,edge blurring,color deviation,and stripe)of image intrinsic properties,we present an adaptive salt-and-pepper denoising method based on a deep residual network.The main idea of this paper is to simplify image denoising into two steps.Firstly,in order to enable the network model to handle different-densities salt-and-pepper noises and improve the robustness of the network model,we remove the high frequency information using adaptive windows as the first step.Secondly,we construct an effective deep residual network model to train a function which can map the pre-processed images to their corresponding ground truths.Qualitative and quantitative experiments show that not only can our method avoid problems such as color distortions and streaks,but also our method outperforms the state-of-the-art learning-based and traditional approaches,in terms of both handling inputs with different levels of noises and revealing high-fidelity image edges.Meanwhile,the performance on BSD300 evaluated in PSNR shows superiority over the competitors.

关 键 词：椒盐噪声 残差网络 图像去噪 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]