基于注意力机制和残差块的真实图像去噪  

Real-world images denoising based on attention mechanism and residual block

作　　者：周联敏 周冬明[1] ZHOU Lian-min;ZHOU Dong-ming(School of Information Science and Engineering,Yunnan University,Kunming 650000,China)

机构地区：[1]云南大学信息学院,云南昆明650000

出　　处：《计算机工程与设计》2023年第5期1451-1458,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(62066047、61365001)。

摘　　要：为有效去除真实图像噪声,提出一种基于注意力机制和残差块的图像去噪算法。采用通道和空间注意力机制相结合的双重注意力模块给不同的特征赋予权重,并与残差块合并于编解码结构;应用增大感受野模块,在保护图像结构的同时增大感受野;整体架构采用密集特征融合保留更多细节。实验结果表明,该算法在SIDD和DND数据集的峰值信噪比分别达到了39.59 dB和39.73 dB,结构相似性分别达到了0.911%和0.953%,具有最佳视觉效果。所提算法在去噪的同时能够保留更多图像细节信息。To effectively remove real-world image noise,an image denoising algorithm based on attention mechanisms and residual blocks was proposed.Dual attention modules combined with channel and spatial attention mechanisms were used to adjust the weights of different features.The module and residual block were combined into an encoding-decoding structure.An enlarging receptive field module was applied to enlarge the receptive field while protecting the image structure.Overall architecture used dense feature fusion to retain more details.Experimental results demonstrate that the peak signal-to-noise ratios of the algorithm in SIDD and DND datasets are 39.59 dB and 39.73 dB,and the structural similarity is 0.911%and 0.953%.It also has the best visual performances.The proposed algorithm can retain more image detail information while denoising.

关 键 词：图像去噪 真实噪声 深度学习 注意力机制 残差块 卷积神经网络 空洞卷积 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]