Reference–Consistent Structural Object Recovery from Tri-Modal SAR–Optical Imagery

Le résumé fourni par la source

Cross-modal remote sensing images of the same scene share stable geometric structures across sensors while exhibiting substantial differences in imaging mechanism, radiometric properties, and noise distribution. Existing methods implicitly rely on this shared structure for registration, translation, and change detection, yet treat it as a latent or embedded quantity rather than an explicit, pixel-domain object that can be independently observed and evaluated. This paper addresses that gap by recovering a common structural object from trimodal inputs. We recover three reference-consistent commonstructure maps and separate them from modality-specific residual components. This is achieved through a shared-reference projection framework, a dual-branch structure-residual design, and a tri-modal local-geometric gating mechanism based on structure tensors. Experiments validate the properties of the recovered object through cross-modal correspondence stability, selective response to object-level change, and structural reorganization under reference substitution, and further test its downstream geometric usability via sparse multi-point correspondences. Results show that the same object is reliably identified across modalities, that genuine changes are isolated from irrelevant variation, and that multi-point correspondences remain sufficiently consistent to support spatial measurement.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.