.png){kind=icon}
.png){kind=icon}
Return
A Diffusion-Guided Task Decomposition Framework for SAR-to-Optical Image Translation
.png){kind=icon}
OA
AI
DOI：10.1109/JSTARS.2025.3623105.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Translating synthetic aperture radar (SAR) images into the optical domain offers a promising solution for enhancing data interpretability and enabling better performance in downstream vision tasks. However, the substantial feature distribution discrepancies between SAR and optical modalities present considerable challenges to existing translation methods, often leading to suboptimal feature alignment and inaccurate translation results. To this end, we propose DGD-S2O, a diffusion-guided task decomposition framework for SAR-to-optical image translation. By leveraging intermediate states in the forward diffusion process, the proposed framework gradually transforms the data distribution from the SAR domain to the optical domain. Specifically, DGD-S2O decomposes the translation task into two synergistic subtasks: modality mapping in the noise domain and image reconstruction through reverse diffusion, which progressively align feature distributions to reduce learning complexity and enhance translation accuracy. Moreover, to ensure semantic consistency between modalities, we introduce a conditional pixel modulator, which adaptively normalizes the injected SAR information at the pixel level during the reconstruction of optical images. In addition, a wavelet-enhanced alignment mechanism is employed to separately process high- and low-frequency components, effectively mitigating low-frequency distribution discrepancies while refining high-frequency details. Extensive experiments on four public benchmarks demonstrate that DGD-S2O not only delivers high-quality image reconstruction but also enhances performance in downstream tasks such as scene classification and land cover segmentation, confirming its effectiveness across diverse scenarios.
Keywords:
Conditional image generation
diffusion model
downstream task enhancement
SAR-to-optical image translation
semantic consistency
synthetic aperture radar (SAR)
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
5.3
Papers:
1.3K
Citations:
3.0W