罗军伟, 张真, 雒芬,等.多尺度注意力交互式图像去噪网络[J].河南理工大学学报（自然科学版）,2023,42(5):144-153.

LUO J W, ZHANG Z, LUO F,et al.Multi-scale attention interactive network for image denoising[J].Journal of Henan Polytechnic University(Natural Science) ,2023,42(5):144-153.

多尺度注意力交互式图像去噪网络

罗军伟, 张真, 雒芬, 乔应旭, 霍占强

河南理工大学 计算机科学与技术学院，河南 焦作 454000

摘要:图像去噪中，针对去噪网络提取图像细节信息不全面和特征利用率低的问题，提出一种基于深度学习的多尺度注意力交互式图像去噪网络（MAINet）。首先，对于浅层像素级特征采用多尺度特征提取块获取丰富的上下文信息和图像纹理特征，以保证图像信息的完整性；然后，引入双路通道注意力机制指导网络获取更具判别性的特征信息，抑制不期望的噪声，从而进一步优化特征信息；最后，利用分类密集残差块的密集连接和成对卷积操作增强模型的交互性，对全局多层次特征进行联合学习，提取更高质量的语义级特征，以提升去噪网络的性能。实验结果表明，在定量和定性评估方面，所提出的去噪网络在合成噪声和真实噪声两种数据集上的去噪效果都有所提升。

关键词:深度学习;图像去噪;多尺度特征提取;双路通道注意力机制;分类密集残差块

doi:10.16186/j.cnki.1673-9787.2021110115

基金项目:国家自然科学基金资助项目（61972134）；河南省高校科技创新团队支持计划项目（19IRTSTHN012）

收稿日期:2021/11/26

修回日期:2022/01/16

出版日期:2023/09/25

Multi-scale attention interactive network for image denoising

LUO Junwei, ZHANG Zhen, LUO Fen, QIAO Yingxu, HUO Zhanqiang

School of Computer Science and Technology，Henan Polytechnic University，Jiaozuo 454000，Henan，China

Abstract:In the image denoising，aiming at the problems of incomplete detail information and low feature utilization of the image extracted by denoising network，a multi-scale attention interactive image denoising network（MAINet）based on deep learning was proposed.Firstly，for shallow pixel-level features，the multi-scale feature extraction block was used to obtain rich context information and image texture features to ensure the integrity of image information.Secondly，a dual-branch attention mechanism was introduced to guide the network to obtain more discriminative feature information，suppress undesired noise，and further optimize the feature information.Finally，the dense connection and the paired convolution operation of the classified dense residual blocks were used to enhance the interaction of the model，and the global multi-level features were jointly learned to extract higher-quality semantic-level features to improve the performance of the denoising network.Experiment results showed that the proposed denoising network had improved the denoising effects on both synthetic noise and real noise datasets in terms of quantitative and qualitative evaluation.

Key words:deep learning;image denoising;multi-scale feature extraction;dual-branch attention mechanism;classified dense residual block

多尺度注意力交互式图像去噪网络_罗军伟.pdf