TY - GEN
T1 - Hybrid Transformer-CNN with Two-Stage Training for SAR Despeckling
AU - Guo, Zhenyu
AU - Hu, Weidong
AU - Zheng, Shichao
AU - Zhang, Yuanyuan
AU - Zhou, Ming
AU - Wang, Qian
AU - Zhang, Ruizhe
N1 - Publisher Copyright:
© 2025 IEEE.
PY - 2025
Y1 - 2025
N2 - Coherent speckle noise in Synthetic Aperture Radar (SAR) imagery severely degrades image fidelity, compromising the accuracy and reliability of subsequent advanced interpretation tasks such as target recognition, scene classification, and semantic segmentation. Convolutional neural networks (CNNs) face a significant domain gap due to their reliance on synthetic data, resulting in insufficient generalization capabilities in real-world scenarios. To address this challenge, we propose a novel denoising framework centered on a hybrid network combining hierarchical Transformers and CNNs. This framework bridges the domain gap through a two-stage 'pre-training-fine-tuning' paradigm and achieves end-to-end optimization via a spatiotemporalfrequency domain joint loss function. Experiments on synthetic and real SAR datasets demonstrate that our approach achieves superior performance across multiple key metrics and visual quality, significantly enhancing despeckling capabilities and detail fidelity.
AB - Coherent speckle noise in Synthetic Aperture Radar (SAR) imagery severely degrades image fidelity, compromising the accuracy and reliability of subsequent advanced interpretation tasks such as target recognition, scene classification, and semantic segmentation. Convolutional neural networks (CNNs) face a significant domain gap due to their reliance on synthetic data, resulting in insufficient generalization capabilities in real-world scenarios. To address this challenge, we propose a novel denoising framework centered on a hybrid network combining hierarchical Transformers and CNNs. This framework bridges the domain gap through a two-stage 'pre-training-fine-tuning' paradigm and achieves end-to-end optimization via a spatiotemporalfrequency domain joint loss function. Experiments on synthetic and real SAR datasets demonstrate that our approach achieves superior performance across multiple key metrics and visual quality, significantly enhancing despeckling capabilities and detail fidelity.
KW - Despeckling
KW - Domain Gap
KW - Multitemporal approach
KW - Synthetic Aperture Radar
KW - Transformers
UR - https://www.scopus.com/pages/publications/105040222599
U2 - 10.1109/CISS67974.2025.11482951
DO - 10.1109/CISS67974.2025.11482951
M3 - Conference contribution
AN - SCOPUS:105040222599
T3 - CISS 2025 - 6th China International SAR Symposium
BT - CISS 2025 - 6th China International SAR Symposium
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 6th China International SAR Symposium, CISS 2025
Y2 - 25 October 2025 through 27 October 2025
ER -