| 周先春,吕梦楠,芮旸,唐彬鑫,杜志亭,陈玉泽.基于注意力机制的双卷积图像去噪网络[J].电子测量与仪器学报,2025,39(2):60-71 |
| 基于注意力机制的双卷积图像去噪网络 |
| Image denoising using dual convolutional neural networkwith attention mechanism |
| |
| DOI： |
| 中文关键词: 图像去噪 卷积神经网络 注意力机制 跳跃连接 多尺度特征提取网络 |
| 英文关键词:image denoising convolutional neural network attention mechanism skip connections multi-scale feature extraction network |
| 基金项目:国家自然科学基金（11202106，61302188）项目资助 |
|
|
| Author | Institution |
| Zhou Xianchun | School of Artificial Intelligence, Nanjing University of Information Science and Technology, Nanjing 210044, China |
| Lyu Mengnan | School of Electronic and Information Engineering, Nanjing University of Information
Science and Technology, Nanjing 210044, China |
| Rui Yang | School of Artificial Intelligence, Nanjing University of Information Science and Technology, Nanjing 210044, China |
| Tang Binxin | School of Electronic and Information Engineering, Nanjing University of Information
Science and Technology, Nanjing 210044, China |
| Du Zhiting | School of Electronic and Information Engineering, Nanjing University of Information
Science and Technology, Nanjing 210044, China |
| Chen Yuze | School of Electronic and Information Engineering, Nanjing University of Information
Science and Technology, Nanjing 210044, China |
|
| 摘要点击次数: 30 |
| 全文下载次数: 62 |
| 中文摘要: |
| 近年来,深度卷积神经网络在图像去噪领域表现出了优越的性能。然而,深度网络结构往往伴随着大量的模型参数,导致训练成本高,推理时间长,限制了其在实际去噪任务中的应用。提出了一种新的基于注意力机制的双卷积图像去噪网络(MA-DFRNet),它由多尺度特征特征提取网络、双卷积神经网络及动态特征精炼注意力机制组成。多尺度特征提取网络通过不同尺度的卷积获取图像特征,提高灵活性。双卷积神经网络上下分支均采用跳跃连接及扩张卷积来增大感受野。动态特征精炼注意力机制增强特征表示的精度和区分能力。这种结构设计不仅扩大了感受野,还更有效地提取和融合图像特征,显著提升去噪效果。研究结果表明,与最先进的模型相比,提出的MA-DFRNet在所有对比的噪声水平下具有更高的峰值信噪比（PSNR）和结构相似性（SSIM）值,PSNR提高了0.2 dB左右,SSIM提高了1%左右,对于噪声水平较高的图像更具鲁棒性,并且在视觉上更好地保留了图像细节,实现去噪和细节保留之间的平衡。 |
| 英文摘要: |
| In recent years, deep convolutional neural networks have shown superior performance in image denoising. However, deep network structures often come with a large number of model parameters, leading to high training costs and long inference times, limiting their practical application in denoising tasks. This paper proposes a new dual convolutional denoising network with skip connections (MA-DFRNet), which achieves an ideal balance between denoising effect and network complexity. The paper presents a novel attention-based dual convolutional image denoising network (MA-DFRNet) that achieves an optimal trade-off between denoising performance and network complexity. MA-DFRNet comprises a multi-scale feature extraction network, dual convolutional neural networks, and a dynamic feature refinement attention mechanism. The multi-scale feature extraction network employs convolutions at various scales to enhance flexibility in capturing image features. The dual convolutional neural networks utilize skip connections and dilated convolutions in both upper and lower branches to expand the receptive field. Furthermore, the dynamic feature refinement attention mechanism enhances the accuracy and discriminability of feature representation. This structural design not only enlarges the receptive field but also effectively extracts and integrates image features, leading to significant improvements in denoising performance. The research findings demonstrate that the proposed MA-DFRNet outperforms state-of-the-art models in terms of PSNR and SSIM values across all levels of noise considered in the comparisons. The PSNR has increased by approximately 0.2 dB, while the SSIM has improved by around 1%. Notably, MA-DFRNet demonstrates greater robustness for images with higher noise levels and better preserves image details visually, effectively balancing denoising and detail retention. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
