多尺度密集交互注意力残差真实图像去噪网络  

Multiscale dense interactive attention residual real image denoising network

作　　者：郭业才[1,2] 胡晓伟 AMITAVE Saha 毛湘南 GUO Yecai;HU Xiaowei;AMITAVE Saha;MAO Xiangnan(College of Electronic and Information Engineering,Nanjing University of Information Science and Technology,Nanjing Jiangsu 210044,China;College of Electronic Information Engineering,Wuxi University,Wuxi Jiangsu 214105,China)

机构地区：[1]南京信息工程大学电子与信息工程学院,江苏南京210044 [2]无锡学院电子信息工程学院,江苏无锡214105

出　　处：《图学学报》2025年第2期279-287,共9页Journal of Graphics

基　　金：国家自然科学基金(61673222)。

摘　　要：针对图像去噪特征提取不全面以及特征利用率低,导致生成图像不够清晰的问题,提出一种多尺度密集交互注意力残差去噪网络(MDIARN)。首先,通过多尺度非对称特征提取模块(MAFM)初步提取浅层信息特征,以确保图像特征的多样性;然后,多尺度级联模块(MSCM)利用多维密集交互残差单元(MDIU)对图像特征进行多维映射,并逐步级联以增强模型之间的信息传递和交互性,充分拟合训练数据;引入双路全局注意力模块(DGAM)对多级特征进行全局联合学习,获取更具有判别性的特征信息;跳跃连接促进结构之间的参数共享,使不同维度的特征充分融合,保证信息的完整性;最后,采用残差学习构建出清晰的去噪图像。结果表明,该算法在真实噪声数据集(DND和SIDD)上峰值信噪比分别为39.80 dB和39.62 dB,结构相似性分别为95.4%和95.8%,均优于主流去噪算法。此外,该算法在低光度场景下应用也能保留更多细节,从而显著提升图像质量。To address the problem that the generated image is not clear enough due to incomplete feature extraction and low feature utilization in image denoising,a multi-scale dense interactive attention residual denoising network(MDIARN)was proposed.First,a multi-scale asymmetric feature extraction module(MAFM)was employed to preliminarily extract shallow information features,ensuring diversity of image features.Then,a multi-scale cascade module(MSCM)utilized multi-dimensional dense interactive residual units(MDIU)to perform multi-dimensional mapping of image features.These units were progressively cascaded to enhance the information transmission and interaction between models,fully fitting the training data.A dual-path global attention module(DGAM)was introduced to conduct global joint learning on multi-level features,acquiring more discriminative feature information.Skip connections were integrated to encourage parameter sharing between structures,enabling full integration of features from different dimensions and preserving the completeness of information.Finally,residual learning was employed to construct a clear denoised image.Experimental results demonstrated that this algorithm achieved peak signal-to-noise ratios of 39.80 dB and 39.62 dB on the real noise datasets(DND and SIDD),respectively,and structural similarities of 95.4% and 95.8%, respectively, outperforming mainstream denoising algorithms. In addition, the proposed algorithm demonstrated excellent performance in low-light environments, preserving more details and significantly enhancing image quality.

关 键 词：图像去噪 多尺度特征提取 多维密集交互 卷积神经网络 注意力 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]