SAR-Conditioned Flow Matching for SAR-to-Optical Image Translation

Jeongwon Ryu, Youngtack Oh, Minseok Seo

TL;DR: By modeling SAR-to-optical translation as a deterministic SAR-conditioned flow, we achieve structure-preserving translation with improved image quality over prior methods.

Abstract: Synthetic Aperture Radar (SAR) imagery enables reliable observation regardless of weather conditions or acquisition time; however, its interpretation is challenging due to speckle noise and complex scattering characteristics that obscure spatial structures. To alleviate this limitation, SAR-to-Optical image translation has been proposed to translate SAR imagery into optical-like images. Existing SAR-to-optical image translation models can struggle to preserve structural consistency, such as object layouts and spatial relationships, as they primarily optimize for distribution matching rather than structure-level fidelity. In this paper, we formulate SAR-to-optical image translation from the perspective of structural consistency, where preserving spatial correspondence between input and output is essential. We introduce a flow-based framework that learns SAR-conditioned transformations defined by ordinary differential equations (ODE), modeling translation as a trajectory that continuously deforms SAR-aligned latent representations into their optical counterparts. By explicitly tracking how each SAR structure evolves along this trajectory, the framework naturally preserves geometric layouts and spatial relationships encoded in SAR imagery. Quantitative and qualitative results demonstrate improved image quality over existing image translation baselines, while preserving structural properties inherent in SAR images.

Submission Number: 27