融合多头注意力机制的图像降噪网络模型  

Image Denoising Network Model Combined with Multi-head Attention Mechanism

作　　者：李玥玥 刘万平[1] 黄东 LI Yueyue;LIU Wanping;HUANG Dong(College of Computer Science and Engineering,Chongqing University of Technology,Chongqing 400054,China;Key Laboratory of Advanced Manufacturing Technology of the Ministry of Education,Guizhou University,Guiyang 550025,China)

机构地区：[1]重庆理工大学计算机科学与工程学院,重庆400054 [2]贵州大学现代制造技术教育部重点实验室,贵阳550025

出　　处：《计算机科学》2023年第S02期314-321,共8页Computer Science

基　　金：重庆市自然科学基金(cstc2021jcyj-msxmX0594);重庆理工大学研究生创新项目资助(gzlcx20223212)。

摘　　要：由于GPU计算的快速发展,深度学习近年来在图像降噪方面得到了应用。大多数深度学习方法都需要无噪声图像作为训练标签,但通常它们很难获得,甚至不可能获得。于是,有学者开始研究使用噪声图像进行降噪网络训练,但其恢复的图像却面临丢失细节信息的问题。受Noise2Noise(N2N)的思想启发,文中使用成对的噪声图像训练神经网络,学习同一范围的同类型噪声之间的分布关系,实现了一种新的降噪网络模型。新开发的模型(MA-UNet)基于经典UNet架构,融合了多头注意力机制(Multi-head Attention)和简易的残差网络,可以更好地挖掘图像的关键信息,掌握特征的全局信息,从而恢复更清晰的图像。与传统算法(CBM3D)和其他方法(如DnCNN和B2U)相比,MA-UNet的性能参数优良。从视觉图像观察来看,所提模型恢复了更清晰的图像细节。与N2N设计的模型相比,在不同噪声幅值下,所提模型在4个经典数据集上的峰值信噪比和结构相似性指数的均值均有显著提高。Due to the rapid development of GPU computing,deep learning has been applied in image denoising recently.Most of the deep learning methods require noise-free images as training labels,but they are usually difficult or even impossible to obtain.Therefore,some scholars begin to study the use of noisy images for noise reduction network training,but the restored image is faced with the problem of losing details.Inspired by the idea of Noise2Noise(N2N),this paper uses pairs of noised images to train the neural network,to learn the distribution relationship between the same type of noise in the same range,and realize a new novel image denoising network model.The newly-developed model(MA-UNet)is based on the classic UNet architecture and combines the multi-head attention mechanism and simple residual network.It can capture the key information of the image,master the glo-bal information of the feature,so as to recover clearer images.Compared with the traditional algorithm CBM3D and other me-thods,such as DnCNN and B2U,MA-UNet has excellent performance in terms of parameters.Through the comparison of visual images,our model restores much clearer image details.Compared with the model designed by N2N,under different noise magnitude,the mean value of the peak signal-to-noise ratio and the structural similarity index of the proposed model on four classical data sets improve significantly.

关 键 词：深度学习 注意力机制 细节信息 图像降噪 全局特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]