[1]郭 聪,杨 敏.基于无参注意力和特征融合的图像去噪算法[J].计算机技术与发展,2023,33(02):50-56.[doi:10. 3969 / j. issn. 1673-629X. 2023. 02. 008]
　GUO Cong,YANG Min.Algorithm of Image Denoising Based on Nonparametric Attention Mechanism and Feature Fusion[J].,2023,33(02):50-56.[doi:10. 3969 / j. issn. 1673-629X. 2023. 02. 008]
点击复制

基于无参注意力和特征融合的图像去噪算法()
分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
33
期数:
2023年02期
页码:
50-56
栏目:
媒体计算
出版日期:
2023-02-10

文章信息/Info

Title:
Algorithm of Image Denoising Based on Nonparametric Attention Mechanism and Feature Fusion
文章编号:
1673-629X(2023)02-0050-07
作者:
郭 聪; 杨 敏
南京邮电大学 自动化学院、人工智能学院,江苏 南京 210023
Author(s):
GUO Cong; YANG Min
School of Automation and School of Artificial Intelligence,Nanjing University of Posts and Telecommunications,Nanjing 210023,China
关键词:
图像去噪; 注意力机制; 特征融合; 图像处理; 卷积神经网络
Keywords:
×image denoising; attention mechanism; feature fusion; image processing; convolution neural network
分类号:
TP391. 4
DOI:
10. 3969 / j. issn. 1673-629X. 2023. 02. 008
摘要:
针对传统图像去噪网络的恢复图像存在纹理和条纹不清晰的问题,提出基于无参注意力机制和特征融合的图像去噪网络 NAFDNet。 该网络包括普通卷积层、注意力特征提取模块和特征融合增强模块。 首先,利用普通卷积层提取的浅层特征作为全局特征。 接着,在注意力特征提取模块中,网络通过混合空洞卷积组和普通卷积相结合,提取特征,针对提取的特征引入无参注意力机制,关注特征图中具有丰富纹理和细节信息。 特征融合增强模块利用两个 1 ×1 卷积,分别学习全局特征图和局部特征图的权重,与对应特征图相乘后相加,获得更为健壮的融合特征。 实验结果表明:NAFDNet 算法在 Set12 测试集上具有较好的客观指标,并且去噪图像具有更清晰的边缘以及纹理特征。
Abstract:
To solve the problem of unclear texture and fringe in the restored image of traditional image denoising network, denoisingnetwork with nonparametric attention mechanism and feature fusion ( NAFDNet ) is proposed, which includes common convolutions,attention feature extracted block and feature fusion enhanced block. Firsly,shallow features extracted from common convolution layer areused as global features. Then,in the attention feature extracted block,the network extracts features through the combination of mixed holeconvolution sets and ordinary convolution,and introduces the non-parametric attention mechanism for the extracted features,focusing onthe feature map with rich texture and detail information. Feature fusion enhanced block uses two 1×1 convolution to learn the weights ofglobal feature graphs and local feature graphs respectively,and then multiply and add them with corresponding feature graphs to obtainmore robust fusion features. The experiment shows that NAFDNet algorithm has a better objective index on Set12 test set, and thedenoised images have clearer edge and texture features.

相似文献/References:

[1]黄伟 王书文.一种基于图像边缘检测的全变分的去噪方法[J].计算机技术与发展,2009,(02):24.
　HUANG Wei,WANG Shu-wen.Total Variation Image Denoising Based on Image Edge Detection[J].,2009,(02):24.
[2]周先国 李开宇.基于提升小波结合DCT变换的图像去噪研究[J].计算机技术与发展,2009,(02):62.
　ZHOU Xian-guo,LI Kai-yu.Image De - noising Research Based on Lifting Wavelet and Discrete Cosine Transform[J].,2009,(02):62.
[3]邢丹俊 王继成.基于提升小波的自适应阈值图像去噪[J].计算机技术与发展,2008,(02):42.
　XING Dan-jun,WPNG Ji-cheng.Adaptive Threshold Based on Lifting Wavelet Transform for Image Denoising[J].,2008,(02):42.
[4]于梅 殷兵 梁栋 王璐.基于边缘检测的Contourlet变换图像去噪[J].计算机技术与发展,2008,(06):104.
　YU Mei,YIN Bing,LIANG Dong,et al.Image Denoising with Contourlet Transform Based on Edge Detection[J].,2008,(02):104.
[5]林德贵 何建农 郑玉燕.基于分水岭的提升小波图像去噪[J].计算机技术与发展,2008,(08):29.
　LIN De-gui,HE Jian-nong,ZHENG Yu-yan.Lifting Wavelet Image De- Noising Based on Watershed[J].,2008,(02):29.
[6]汪继文 林胜华 沈玉峰 邱剑锋.一种基于各向异性扩散的图像处理方法[J].计算机技术与发展,2008,(08):98.
　WANG Ji-wen,LIN Sheng-hua,SHEN Yu-feng,et al.An Approach for Image Restoration Based on Anisotropic Diffusion[J].,2008,(02):98.
[7]李瑶 董瑞 何韬 梁栋.一种基于Contourlet变换的图像去噪方法[J].计算机技术与发展,2007,(03):81.
　LI Yao,DONG Rui,HE Tao,et al.A Method for Image Denoising Based on Contourlet Transform[J].,2007,(02):81.
[8]王国权 刘亮.FoE模型的训练方法研究[J].计算机技术与发展,2010,(12):86.
　WANG Guo-quan,LIU Liang.A Study on Training Methods of FoE Model[J].,2010,(02):86.
[9]方洁.基于非抽样Contourlet变换的最佳软阈值图像去噪[J].计算机技术与发展,2011,(02):102.
　FANG Jie.Image Denoising Based on Nonsubsampled Contourlet Transform and Best Soft Threshold[J].,2011,(02):102.
[10]柯丹丹 蔡光程 曹倩倩.基于形态学算子的各向异性扩散去噪方法[J].计算机技术与发展,2012,(04):81.
　KE Dan-dan,CAI Guang-cheng,CAO Qian-qian.An Anisotropic Diffusion Denoising Method Based on Morphological Operator[J].,2012,(02):81.

更新日期/Last Update: 2023-02-10