基于双分支动态特征增强的SAR图像去噪

doi:10.3969/j.issn.1000-1565.2026.04.011

摘要/Abstract

摘要： 针对现有去噪算法常常忽视图像低对比度区域的问题,本文提出一种用于散斑噪声抑制的双分支网络——双分支动态特征增强网络(dual-branch dynamic feature enhancement network,DDFE-Net).DDFE-Net包含2个基于U形网络的分支,分别是低对比度区域特征增强(feature enhancement in low contrast areas,FELCA)分支和小波域特征增强(wavelet domain feature enhancement,WDFE)分支.FELCA分支首先通过动态特征提取(dynamic feature extraction,DFE)模块对图像的互补特征进行提取,并采用动态多头Transformer(dynamic multi-head attention Transformer,DMFormer)捕获全局特征,从而增强了低对比度区域的特征.WDFE分支首先利用残差模块对含噪声的输入图像进行底层特征编码,再通过动态小波域关联器(dynamic wavelet domain correlator,DWDC)模块建立图像小波不同频率之间的对应关系,进一步增加图像的空间连续性.最终,DDFE-Net将2个分支所获得的特征进行融合,重建出高质量去噪图像.实验结果表明:DDFE-Net在噪声抑制、边缘保护和计算效率方面均优于现有方法,尤其在处理复杂噪声时表现出更强的鲁棒性.

关键词: SAR图像去噪, 深度学习, 动态特征提取, 动态小波域关联器

Abstract: Aiming at the problem that the existing denoising algorithms often ignore the low contrast region of the image, we propose a dual-branch network for speckle noise suppression, namely, the dual-branch- 引用格式:冯忠居,于明威,张聪,等.冲刷场地地震波形对大直径变截面单桩时程响应的影响［J］.河北大学学报(自然科学版),2026,46(4):337-348.引用格式:刘明,周永露,赵宇航,等.基于双分支动态特征增强的SAR图像去噪［J］.河北大学学报(自然科学版),2026,46(4):437-448.DOI:10.3969/j.issn.1000-1565.2026.04.011基于双分支动态特征增强的SAR图像去噪刘明^1,2,周永露^1,3,赵宇航^1,3,刘帅奇^1,3,赵淑欢^1,3(1.河北大学电子信息工程学院,河北保定 071002;2.河北大学信息技术中心,河北保定 071002;3.河北省机器视觉技术创新中心,河北保定 071002)摘要:针对现有去噪算法常常忽视图像低对比度区域的问题,本文提出一种用于散斑噪声抑制的双分支网络——双分支动态特征增强网络(dual-branch dynamic feature enhancement network,DDFE-Net).DDFE-Net包含2个基于U形网络的分支,分别是低对比度区域特征增强(feature enhancement in low contrast areas,FELCA)分支和小波域特征增强(wavelet domain feature enhancement,WDFE)分支.FELCA分支首先通过动态特征提取(dynamic feature extraction,DFE)模块对图像的互补特征进行提取,并采用动态多头Transformer(dynamic multi-head attention Transformer,DMFormer)捕获全局特征,从而增强了低对比度区域的特征.WDFE分支首先利用残差模块对含噪声的输入图像进行底层特征编码,再通过动态小波域关联器(dynamic wavelet domain correlator,DWDC)模块建立图像小波不同频率之间的对应关系,进一步增加图像的空间连续性.最终,DDFE-Net将2个分支所获得的特征进行融合,重建出高质量去噪图像.实验结果表明:DDFE-Net在噪声抑制、边缘保护和计算效率方面均优于现有方法,尤其在处理复杂噪声时表现出更强的鲁棒性.关键词:SAR图像去噪;深度学习;动态特征提取;动态小波域关联器中图分类号:TN95 文献标志码:A 文章编号:1000-1565(2026)04-0437-12DOI:10.3969/j.issn.1000-1565.2026.04.011Dual-branch dynamic feature enhancement network for SAR image denoisingLIU Ming^1,2, ZHOU Yonglu^1,3, ZHAO Yuhang^1,3, LIU Shuaiqi^1,3, ZHAO Shuhuan^1,3(1. College of Electronic and Informational Engineering, Hebei University, Baoding 071002, China;2. Information Technology Center, Hebei University, Baoding 071002, China;3. Machine Vision Technology Innovation Center of Hebei Province, Baoding 071002, China)Abstract: Aiming at the problem that the existing denoising algorithms often ignore the low contrast region of the image, we propose a dual-branch network for speckle noise suppression, namely, the dual-branch- 收稿日期:2025-08-28;修回日期:2026-01-13 基金项目:国家自然科学基金项目(62172139);中央引导地方科技发展基金项目(246Z0104G);河北省高等学校科学技术研究项目(BJ2026066) 第一作者:刘明(1985—),女,河北大学实验师,主要从事算法优化、多维信号处理、图像处理方向研究. E-mail:mliu_hbu@163.com 通信作者:赵淑欢(1987—),女,河北大学讲师,主要从事图像处理和信号处理方向研究. E-mail:zsh_hbu@163.com 第4期刘明等:基于双分支动态特征增强的SAR图像去噪河北大学学报(自然科学版) 第46卷dynamic feature enhancement network(DDFE-Net). DDFE-Net contains two branches based on U-shaped network, namely feature enhancement in low contrast areas(FELCA)branch and wavelet domain feature enhancement(WDFE)branch. The FELCA branch first extracts the complementary features of the image through the dynamic feature extraction(DFE)module, and uses the dynamic multi-head attention Transformer(DMFormer)to capture the global features, to enhance the features of the low contrast region. The WDFE branch first uses the residual module to encode the underlying features of the noisy input image, and then establishes the corresponding relationship between image frequencies in wavelet domain through the dynamic wavelet domain correlator(DWDC)to further increase the spatial continuity of the image. Finally, DDFE-Net fuse the features obtained from the two branches, and then reconstructed a high-quality denoised image. Experimental results show that DDFE-Net is superior to existing methods in noise suppression, edge protection and computational efficiency, especially in dealing with complex noise.

Key words: SAR image denoising, deep learning, dynamic feature extraction, dynamic wavelet domain correlator

中图分类号:

TN95

刘明,周永露,赵宇航,刘帅奇,赵淑欢. 基于双分支动态特征增强的SAR图像去噪[J]. 河北大学学报(自然科学版), 2026, 46(4): 437-448.

LIU Ming, ZHOU Yonglu, ZHAO Yuhang, LIU Shuaiqi, ZHAO Shuhuan. Dual-branch dynamic feature enhancement network for SAR image denoising[J]. Journal of Hebei University(Natural Science Edition), 2026, 46(4): 437-448.

参考文献

[1] 刘帅奇,扈琪,刘彤,等.合成孔径雷达图像去噪算法研究综述[J].兵器装备工程学报, 2018, 39(12): 106-112. DOI:10.11809/bqzbgcxb2018.12.022.
[2] 雷钰,刘帅奇,张璐瑶,等.基于深度学习的合成孔径雷达图像去噪综述[J].兵器装备工程学报, 2022, 43(11): 71-80. DOI:10.11809/bqzbgcxb2022.11.010.
[3] Parrilli S, Poderico M, Angelino C V, et al. A nonlocal SAR image denoising algorithm based on LLMMSE wavelet shrinkage[J]. IEEE Trans Geosci Remote Sens, 2012, 50(2): 606-616. DOI:10.1109/TGRS.2011.2161586.
[4] Kevala V D, Lal S. NBDNet: A deep learning algorithm for despeckling of SAR data[J]. SN Comput Sci, 2024, 5(7): 897. DOI:10.1007/s42979-024-03274-6.
[5] Xiao S Y, Zhang S S, Huang L B, et al. Trans-NLM network for SAR image despeckling[J]. IEEE Trans Geosci Remote Sens, 2024, 62: 5211912. DOI:10.1109/TGRS.2024.3397325.
[6] Liu S Q, Lei Y, Zhang L Y, et al. MRDDANet: A multiscale residual dense dual attention network for SAR image denoising[J]. IEEE Trans Geosci Remote Sens, 2022, 60: 5214213. DOI:10.1109/TGRS.2021.3106764.
[7] Liu S Q, Zhang L Y, Tian S K, et al. MFAENet: A multiscale feature adaptive enhancement network for SAR image despeckling[J]. IEEE J Sel Top Appl Earth Obs Remote Sens, 2023, 16: 10420-10433. DOI:10.1109/JSTARS.2023.3327332.
[8] Wang X, Wu Y X, Shi C T, et al. ANED-net: Adaptive noise estimation and despeckling network for SAR image[J]. IEEE J Sel Top Appl Earth Obs Remote Sens, 2024, 17: 4036-4051. DOI:10.1109/JSTARS.2024.3355220.
[9] Li J F, Wen Y, He L H. SCConv: Spatial and channel reconstruction convolution for feature redundancy[C] //2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR). June 17-24, 2023, Vancouver, BC, Canada. IEEE, 2023: 6153-6162. DOI:10.1109/CVPR52729.2023.00596.
[10] Dosovitskiy A, Beyer L, Kolesnikov A, et al. An image is worth 16x16 words: Transformers for image recognition at scale[EB/OL]. 2020: arXiv: 2010.11929. https://doi.org/10.48550/arXiv.2010.11929.
[11] Hou Q B, Zhou D Q, Feng J S. Coordinate attention for efficient mobile network design[C] //2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR). June 20-25, 2021. Nashville, TN, USA. IEEE, 2021: 13708-13717. DOI:10.1109/cvpr46437.2021.01350.
[12] Yang Y, Newsam S. Bag-of-visual-words and spatial extensions for land-use classification[C] //Proceedings of the 18th SIGSPATIAL International Conference on Advances in Geographic Information Systems. San Jose California. ACM, 2010: 270-279. DOI:10.1145/1869790.1869829.
[13] Dabhi S, Soni K, Patel U, et al. Virtual SAR: A synthetic dataset for deep learning based speckle noise reduction algorithms[EB/OL]. 2020: arXiv: 2004.11021. https://doi.org/10.48550/arXiv.2004.11021.
[14] Zhang T W, Zhang X L, Ke X, et al. LS-SSDD-v1.0: A deep learning dataset dedicated to small ship detection from large-scale sentinel-1 SAR images[J]. Remote Sens, 2020, 12(18): 2997. DOI:10.3390/rs12182997.
[15] Liu S Q, Gao L L, Lei Y, et al. SAR speckle removal using hybrid frequency modulations[J]. IEEE Trans Geosci Remote Sens, 2021, 59(5): 3956-3966. DOI:10.1109/TGRS.2020.3014130.
[16] Liu S Q, Liu T, Gao L L, et al. Convolutional neural network and guided filtering for SAR image denoising[J]. Remote Sens, 2019, 11(6): 702. DOI:10.3390/rs11060702.
[17] Thakur R K, Maji S K. SIFSDNet: Sharp image feature based SAR denoising network[C] //IGARSS 2022 - 2022 IEEE International Geoscience and Remote Sensing Symposium. July 17-22, 2022, Kuala Lumpur, Malaysia. IEEE, 2022: 3428-3431. DOI:10.1109/IGARSS46834.2022.9883415.
[18] Thakur R K, Maji S K. AGSDNet: attention and gradient-based SAR denoising network[J]. IEEE Geosci Remote Sens Lett, 2022, 19: 4506805. DOI:10.1109/LGRS.2022.3166565.
[19] Vitale S, Ferraioli G, Frery A C, et al. SAR despeckling using multiobjective neural network trained with generic statistical samples[J]. IEEE Trans Geosci Remote Sensing, 2023, 61: 1-12. DOI:10.1109/tgrs.2023.3314857.
[20] Guo Y, Lu Y X, Liu R W, et al. Blind image despeckling using a multiscale attention-guided neural network[J]. IEEE Trans Artif Intell, 2024, 5(1): 205-216. DOI:10.1109/TAI.2023.3235342.
[21] Liu S Q, Zhang L Y, Tian S K, et al. MFAENet: A multiscale feature adaptive enhancement network for SAR image despeckling[J]. IEEE J Sel Top Appl Earth Obs Remote Sens, 2023, 16: 10420-10433. DOI:10.1109/JSTARS.2023.3327332.
[22] Perera M V, Bandara W G C, Valanarasu J M J, et al. Transformer-based SAR image despeckling[C] //IGARSS 2022—2022 IEEE International Geoscience and Remote Sensing Symposium. July 17-22, 2022, Kuala Lumpur, Malaysia. IEEE, 2022: 751-754. DOI:10.1109/IGARSS46834.2022.9884596. (