WACV 2026 超分辨率（super-resolution）方向上接收论文总结

WACV 2026目录WACV 2026图像超分参考图像 / 双摄超分任意尺度 / 轻量级超分遥感 / 卫星图像超分红外图像超分其他 / 相关任务总结参考资料WACV 2026 会议于 2026 年 3 月 6 日至 10 日在美国亚利桑那州图森Tucson举行主会时间为 3 月 8 日至 10 日。本文基于 WACV 2026 Open Access 的主会论文列表汇总标题或内容与超分辨率Super-Resolution, SR相关的论文。现将超分辨率方向上接收的论文汇总如下。标题完全不含 Super-Resolution / SR、但摘要或正文明确将超分辨率作为核心任务、评测任务或关键模块的论文统一放入“其他 / 相关任务”中。图像超分参考图像 / 双摄超分DM3Net: Dual-Camera Super-Resolution via Domain Modulation and Multi-scale MatchingPaper: https://arxiv.org/abs/2506.06993Keywords: Dual-Camera Super-Resolution, Reference-based SR, Domain Modulation, Multi-scale Matching, Key PruningFeatures: 面向手机等多摄设备使用长焦图像作为参考增强广角低分辨率图像通过域调制缓解 HR 域与退化域差异通过多尺度 patch 匹配传递高频结构Key Pruning 降低匹配显存和推理时间Team: Waseda University, Zhejiang University任意尺度 / 轻量级超分A Fast, Simple, and Flexible Scale Informative Feature Transform Module for Arbitrary Scale Image Super-ResolutionPaper: openaccessCode: https://github.com/aupendu/FTSRKeywords: Arbitrary-Scale SR, Scale Informative Feature Transform, Fully Convolutional UpscalingFeatures: 提出可插入其他 SR 网络的全卷积任意尺度上采样模块相比隐式表示方法参数更少、显存和推理开销更低还扩展到单应变换下的超分辨率Team: Dolby Laboratories, Indian Institute of Technology KharagpurGRAPE (Gaussian Rendering for Accelerated Pixel Enhancement) Brings Fast and Lightweight Arbitrary Super-ResolutionPaper: openaccessCode: https://github.com/mulkkog/GRAPEKeywords: Arbitrary-Scale SR, 2D Gaussian Splatting, Lightweight SR, Differentiable RasterizationFeatures: 用单个 point-wise layer 预测各向异性 Gaussian 参数再通过可微光栅化一次渲染 HR 图像模型约 1.56M 参数在 Urban100 x4 上报告 69.33 FPS适合资源受限和交互式场景Team: Korea University遥感 / 卫星图像超分Segmentation-Aware Latent Diffusion for Satellite Image Super-Resolution: Enabling Smallholder Farm Boundary DelineationPaper: https://arxiv.org/abs/2511.14481Keywords: Satellite Image SR, Reference-based SR, Latent Diffusion, Segmentation-aware SR, Geo-spatial Foundation ModelsFeatures: 提出 SEED-SR不在像素空间直接恢复 HR 图像而是在面向农田边界分割的 latent space 中做超分利用多源多光谱低分辨率时序影像、历史高分参考影像和地理空间基础模型实现最高 20x 的任务感知 SRTeam: Google DeepMind, Google, Google Research红外图像超分RPT-SR: Regional Prior attention Transformer for infrared image Super-ResolutionPaper: https://arxiv.org/abs/2602.15490Code: https://github.com/Yonsei-STL/RPT-SRBlog: arXiv每日学术速递RPT-SRKeywords: Infrared Image Super-Resolution, Regional Prior Attention, Transformer, LWIR, SWIRFeatures: 针对监控、自动驾驶等固定或近静态视角红外场景引入可学习 regional prior token 作为场景结构记忆与 local token 融合调制重建同时覆盖 LWIR 和 SWIR 数据集Team: Yonsei University, BK21 Graduate Program in Intelligent Semiconductor Technology其他 / 相关任务这一部分收录标题不直接包含 Super-Resolution / SR、但论文摘要或正文中明确将超分辨率作为核心任务、评测任务或关键模块的工作。Flood-LDM: Generalizable Latent Diffusion Models for rapid and accurate zero-shot High-Resolution Flood MappingPaper: https://arxiv.org/abs/2511.14033Code: https://github.com/neosunhan/flood-diffKeywords: Flood Map Super-Resolution, Latent Diffusion, Physics-informed Conditioning, Zero-shot GeneralizationFeatures: 将粗网格洪水图超分到接近细网格水动力模型的精度并用 DEM 作为物理先验条件强调跨流域泛化和实时洪水风险管理Team: National University of Singapore, University of Melbourne, University of Sydney, Delft University of TechnologySuperRivolution: Fine-Scale Rivers from Coarse Temporal Satellite ImageryPaper: https://arxiv.org/abs/2511.09597Keywords: Temporal Satellite Imagery, River Segmentation Resolution, Multi-frame SR, Remote SensingFeatures: 构建 9,810 张低分辨率 Sentinel-2 时序影像与高分辨率河流标签配对数据集比较输入上采样、输出上采样和预训练 SR 模型等策略用低分辨率时序数据提升河流分割和河宽估计Team: University of Massachusetts AmherstRestora-Flow: Mask-Guided Image Restoration with Flow MatchingPaper: https://arxiv.org/abs/2511.20152Code: https://github.com/imigraz/Restora-FlowBlog: SwiftScholar 解读Keywords: Flow Matching, Training-free Restoration, Mask-guided Sampling, Super-Resolution, Inpainting, DenoisingFeatures: 训练无关的 flow matching 图像恢复方法用 degradation mask 和轨迹校正保持与退化输入一致在自然图像和医学图像上评测 inpainting、super-resolution、denoising 等任务Team: Medical University of Graz, Graz University of Technology, University of ZurichEquivariant Sampling for Improving Diffusion Model-based Image RestorationPaper: https://arxiv.org/abs/2511.09965Code: https://github.com/FouierL/EquSKeywords: Diffusion Model-based Image Restoration, Equivariant Sampling, Zero-shot Restoration, Bicubic SRFeatures: 提出 EquS / EquS通过双采样轨迹引入等变约束并用 timestep-aware schedule 强化确定性采样阶段在去模糊、压缩感知、inpainting、4x bicubic SR、colorization 等任务上作为通用 DMIR 增强模块Team: University of Science and Technology of China, MIRACLE Center, Institute of Computing Technology CASNeural Geometry Image-Based Representations with Optimal Transport (OT)Paper: https://arxiv.org/abs/2511.18679Blog: Emergent Mind Topic, Slideshare 解读Keywords: Geometry Image Super-Resolution, 3D Mesh Compression, Optimal Transport, Continuous LoDFeatures: 将不规则 3D mesh 转换为规则 geometry image mipmap用 CNN/PixelShuffle 做 geometry image 超分单次前向恢复高质量网格通过最优传输缓解平坦区域过采样和细节区域欠采样支持连续 LoDTeam: Futurewei Technologies, Stony Brook University, University of Southern California, Massachusetts Institute of Technology, George Mason University总结从 WACV 2026 的超分辨率相关论文来看本届 SR 工作数量不算多但应用指向非常明确任意尺度与效率仍是主线FTSR 和 GRAPE 都聚焦任意尺度超分并把部署成本、显存、FPS 作为核心卖点。真实应用场景更加具体DM3Net 面向手机双摄RPT-SR 面向固定视角红外监控SEED-SR、Flood-LDM、SuperRivolution 则把 SR 放入遥感、农业、水文等下游应用中。扩散 / Flow Matching 作为恢复先验继续扩展SEED-SR、Flood-LDM、Restora-Flow、EquS 都体现了生成式先验在 SR 或通用图像恢复任务中的持续渗透。任务感知 SR 趋势明显SEED-SR 不追求像素空间“好看”的超分而是直接优化农田边界分割 latentSuperRivolution 也以河流分割和河宽估计为最终目标。SR 概念向非传统图像域扩展Neural Geometry Image-Based Representations 将 3D mesh 表示转为 geometry image 后进行超分说明 SR 思路正被迁移到 3D 表示压缩、重建与 LoD 控制中。总体而言WACV 2026 的超分辨率论文更偏应用驱动和系统落地不是单纯追求通用图像 SR 指标而是围绕移动摄影、红外感知、遥感监测、医学/自然图像恢复和 3D 几何表示强调效率、泛化、下游任务收益与真实数据适配。参考资料WACV 2026 Official DatesWACV 2026 Open AccessWACV 2026 Accepted PapersAwesome-Super-Resolution 论文列表注文档部分内容由 AI 生成论文链接、代码链接与团队单位基于公开页面检索整理。