← 研踪 PaperTrailarXiv:2610.13547 · v2
MOSAIC: Learning Object-Centric Structural Representations for Remote Sensing Image Super-Resolution
首次提交 2026-10-09 21:03
版本更新 2026-11-17 23:19
雷达观测 2026-11-18 16:38
摘要
Remote sensing image super-resolution (RSISR) requires faithfully reconstructing small foreground objects, multi-scale structures, and repetitive spatial patterns from degraded observations. Despite recent advances in generative super-resolution, existing methods often rely on dense low-resolution (LR) conditioning, which can entangle salient foreground structures with dominant backgrounds and insufficiently exploit complementary structural evidence across scales and samples. To address these limitations, we propose MOSAIC, an object-centric structural representation learning framework that complements dense spatial conditioning with compact structural guidance for generative RSISR. Specifically, MOSAIC extracts object-centric structural slots to emphasize salient foreground evidence, adaptively integrates stable cross-scale structures with fine-grained details to address substantial scale variation, and exploits recurring spatial patterns through image-guided retrieval from a shared prototype memory. The resulting representations combine image-specific evidence with shared structural priors to provide scene-level guidance for faithful reconstruction. Extensive experiments on UCMerced, AID, and RSSCN7 demonstrate consistent PSNR improvements across and settings, with gains of 2.07, 4.54, and 1.49 dB over DiT4SR at , respectively, alongside favorable LPIPS results and improved structural preservation. The code will be publicly released.
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v2 · 2026-11-17 23:19 · 最新观测版本
MOSAIC: Learning Object-Centric Structural Representations for Remote Sensing Image Super-Resolution
Remote sensing image super-resolution (RSISR) requires faithfully reconstructing small foreground objects, multi-scale structures, and repetitive spatial patterns from degraded observations. Despite recent advances in generative super-resolution, existing methods often rely on dense low-resolution (LR) conditioning, which can entangle salient foreground structures with dominant backgrounds and insufficiently exploit complementary structural evidence across scales and samples. To address these limitations, we propose MOSAIC, an object-centric structural representation learning framework that complements dense spatial conditioning with compact structural guidance for generative RSISR. Specifically, MOSAIC extracts object-centric structural slots to emphasize salient foreground evidence, adaptively integrates stable cross-scale structures with fine-grained details to address substantial scale variation, and exploits recurring spatial patterns through image-guided retrieval from a shared prototype memory. The resulting representations combine image-specific evidence with shared structural priors to provide scene-level guidance for faithful reconstruction. Extensive experiments on UCMerced, AID, and RSSCN7 demonstrate consistent PSNR improvements across and settings, with gains of 2.07, 4.54, and 1.49 dB over DiT4SR at , respectively, alongside favorable LPIPS results and improved structural preservation. The code will be publicly released.
查看此版本解读 →查看此版本原文 ↗