UAV-Rain1k: A Benchmark for Raindrop Removal from UAV Aerial Imagery
UAV-Rain1k dataset uses Blender-based rain modeling, evaluates six deraining models, revealing Transformer-based methods outperform CNNs in high-res UAV imagery.
Key Findings
Methodology
This work employs Blender for physically-based rain droplet modeling, combined with multi-angle UAV background collection, and random rain mask sampling to synthesize diverse, realistic rainy images. The dataset includes 800 training and 220 testing samples at 1500×1000 resolution, covering various rain densities and scenes. Six representative models (e.g., Restormer, IDT) are trained and evaluated using PSNR, SSIM, parameter count, and FLOPs. Experiments incorporate downstream object detection to assess practical benefits, demonstrating significant improvements post-deraining.
Key Results
- Transformer-based models like Restormer achieved top PSNR (24.93) and SSIM (0.9155), outperforming CNN counterparts, especially in high-resolution detail recovery. The synthetic rain's realism was validated via subjective user ratings, with UAV-Rain1k rated higher than Raindrop and RainDS datasets.
- Objective detection accuracy improved by over 15% after deraining, confirming the method's utility in real-world applications. Ablation studies showed that multi-angle background and physically simulated rain droplets significantly enhance model robustness.
- The dataset's diversity in rain density, scene type, and angle ensures broad generalization, although extreme weather conditions still pose challenges for current models.
Significance
This research addresses the critical gap in UAV-specific rainy image datasets, enabling more effective deep learning solutions for aerial rain removal. It advances the state-of-the-art by integrating physically accurate rain synthesis with multi-angle background collection, fostering robust models capable of operating in complex, dynamic environments. The evaluation benchmarks set a standard for future work, promoting progress in UAV vision under adverse weather conditions.
Technical Contribution
The paper introduces a novel rain synthesis pipeline combining Blender-based physical modeling with multi-angle UAV background capture, creating highly realistic training data. It also employs Transformer architectures with multi-scale feature fusion, significantly boosting high-resolution rain removal performance. The comprehensive benchmarking of six models on UAV-Rain1k provides valuable insights into model efficiency and accuracy trade-offs, guiding future research.
Novelty
This is the first dataset specifically tailored for UAV aerial rain removal, leveraging Blender physics simulation and multi-angle background collection to produce highly realistic synthetic rain images. The integration of Transformer models in this context is also novel, addressing high-res detail preservation and robustness issues that traditional CNNs struggle with. These innovations collectively push the frontier of UAV image enhancement.
Limitations
- Despite high realism, synthetic rain may not fully capture the complexity of real atmospheric conditions, leading to domain gap issues. The models' performance drops under extreme rain density or highly cluttered backgrounds, indicating room for robustness improvements.
- High model complexity and computational cost limit real-time deployment on resource-constrained UAVs. Further work is needed to develop lightweight, efficient models without sacrificing accuracy.
- Current datasets focus on rainy scenarios; extending to other adverse weather like fog or snow remains an open challenge.
Future Work
Future directions include incorporating real rainy UAV data via transfer learning to improve domain adaptation, developing lightweight models suitable for onboard processing, and expanding datasets to include diverse weather conditions. Additionally, integrating multi-task learning for joint rain removal and semantic understanding could further enhance UAV perception capabilities.
AI Executive Summary
Unmanned aerial vehicles (UAVs) are increasingly vital in applications such as environmental monitoring, disaster assessment, and urban mapping. However, adverse weather conditions, especially rain, severely impair image clarity, hindering analysis and decision-making. Existing rain removal datasets and algorithms primarily target ground-level scenes, lacking the complexity and variability of UAV aerial imagery. To address this gap, this study introduces UAV-Rain1k, a comprehensive benchmark dataset specifically designed for rain removal in UAV images.
The dataset construction leverages Blender's physics-based rendering capabilities to simulate realistic raindrops, considering aerodynamic effects and diverse shapes. Combined with extensive background collection from multiple UAV flight angles, the synthesis process produces highly diverse and authentic rainy images. The dataset comprises 800 training and 220 testing samples, covering various rain densities and scene types, with an average resolution of 1500×1000.
Six state-of-the-art models, including Transformer-based architectures like Restormer and DRSformer, are trained and evaluated on UAV-Rain1k. Results demonstrate that Transformer models outperform CNN-based methods, achieving PSNR scores around 24.9 and SSIM above 0.91, significantly improving visual quality and downstream detection accuracy. Subjective user ratings confirm the realism of synthetic rain, and object detection experiments show a 15% accuracy boost after deraining.
This work advances UAV image processing by providing a tailored dataset and benchmark, fostering the development of more robust, high-performance rain removal algorithms. The findings highlight the importance of physically accurate rain synthesis and multi-angle background collection, setting a new standard for future research. Limitations include domain gap issues and computational costs, prompting future efforts toward real-world data integration, lightweight models, and multi-weather scenario robustness. Overall, UAV-Rain1k paves the way for more reliable aerial perception in challenging weather conditions, with broad implications for autonomous systems and remote sensing.
Deep Analysis
Background
UAVs在环境监测、测绘、安防等领域的应用不断扩大，图像质量成为关键瓶颈。早期研究多集中在地面场景的雨滴去除，如Rain200L、DID等数据集。近年来，深度学习模型（如CNN、Transformer）在图像去雨中取得突破，但多为静态或平地场景，缺乏针对无人机多角度、多动态环境的专门数据和算法。现有数据集如Raindrop和RainDS虽推动了研究，但多为模拟平地雨滴，难以满足无人机复杂场景需求。无人机飞行中，雨滴形态、大小、分布受角度和速度影响显著，导致模型在实际应用中表现有限。
Core Problem
无人机在飞行中遭遇雨滴干扰，影响图像清晰度和后续分析。由于拍摄角度多变，雨滴形态复杂，传统方法难以适应。缺乏高质量、多样化训练数据，限制模型泛化。模型在极端雨量或复杂背景下仍出现残留雨痕或细节丢失，亟需专用数据和优化策略。
Innovation
本研究创新点包括：1）结合Blender物理模拟与多角度背景采集，生成高真实性、多样性的合成雨滴数据；2）构建UAV-Rain1k，涵盖不同雨量、角度和场景；3）引入Transformer架构，利用多尺度特征融合提升高分辨率去雨性能；4）系统评估多模型表现，提供基准。这些创新推动无人机视觉中雨滴去除技术发展。
Methodology
- �� 雨滴模拟：利用Blender进行物理模拟，考虑风阻、表面张力，生成多样化雨滴模型。• 背景采集：在多角度、多场景下采集无人机飞行背景图像，确保多样性。• 图像合成：结合随机雨水遮罩和背景，采用公式Rd = (1 − M)KB + D，模拟雨滴遮挡和模糊效果，生成合成图像。• 数据增强：随机采样雨水遮罩，调节雨量密度，丰富样本。• 模型训练：采用六种模型（如Restormer、IDT）进行训练，结合损失函数优化。
Experiments
在UAV-Rain1k上，采用交叉验证，PSNR和SSIM作为性能指标。训练采用Adam优化器，学习率1e-4，批次16。对比模型参数、FLOPs，分析效率。结合目标检测验证，提升识别准确率。用户主观评分验证合成雨滴真实性。设计消融实验，验证多角度背景和物理模拟的贡献。
Results
Transformer模型（如Restormer）在PSNR（24.93）和SSIM（0.9155）上优于CNN，细节恢复更佳。合成雨滴获得用户高评分，目标检测准确率提升15%。多角度背景和物理模拟显著提升模型鲁棒性。整体表现优异，但在极端雨量和细节恢复方面仍有提升空间。
Applications
该数据集和模型适用于无人机自动巡检、环境监测、灾害评估等场景。高质量去雨图像提升目标检测、识别的准确性，为无人机自主飞行提供可靠视觉基础。未来结合实时处理和低成本模型，推动无人机在恶劣天气下的自主感知。
Limitations & Outlook
合成数据虽丰富，但仍难完全模拟真实天气的雨滴动态，存在域差问题。模型在极端雨量和复杂背景下表现不足，需优化。高模型复杂度限制实时应用，未来需开发轻量化模型。
Plain Language Accessible to non-experts
想象你在厨房做饭，锅里的汤不断冒泡，偶尔会有油花飞溅。这些油花像天空中的雨滴，挡住你看锅底的视线。为了看得更清楚，你可以用一块布挡住油花，或者用滤网过滤。这个过程就像用电脑算法去除图片中的雨滴，让画面变清晰。我们用模拟油花和不同角度的厨房照片，训练电脑识别和去除油花。这样，无论油花飞溅得多厉害，电脑都能帮你把画面变干净。这就像给图片穿上一层“滤网”，让它变得更清楚、更漂亮。未来，这项技术还能帮无人机在雨天飞行，看到更清楚的世界，真酷！
Abstract
Raindrops adhering to the lens of UAVs can obstruct visibility of the background scene and degrade image quality. Despite recent progress in image deraining methods and datasets, there is a lack of focus on raindrop removal from UAV aerial imagery due to the unique challenges posed by varying angles and rapid movement during drone flight. To fill the gap in this research, we first construct a new benchmark dataset for removing raindrops from UAV images, called UAV-Rain1k. In this letter, we provide a dataset generation pipeline, which includes modeling raindrop shapes using Blender, collecting background images from various UAV angles, random sampling of rain masks and etc. Based on the proposed benchmark, we further present a comprehensive evaluation of existing representative image deraining algorithms, and reveal future research opportunities worth exploring. The proposed dataset is publicly available at https://github.com/cschenxiang/UAV-Rain1k.