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 paper, 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.
Introduction
Unmanned Aerial Vehicles (UAVs) are increasingly utilized for surveillance, mapping, and disaster monitoring. However, their operational efficiency is significantly hampered by adverse weather conditions, particularly rain. Raindrops adhering to a drone's camera lens create complex occlusions, color distortions, and spatial information loss. While image deraining is a well-studied field, existing benchmarks primarily focus on street-level views typical of autonomous driving. Drones operate at varying altitudes and steep downward angles, resulting in raindrop patterns and background textures that differ substantially from traditional datasets. UAV-Rain1k addresses this gap by providing the first specialized benchmark for aerial raindrop removal.
Dataset Generation Pipeline
To create a realistic synthetic dataset, the authors developed a sophisticated pipeline leveraging physical modeling and diverse background collection.
- Physical Modeling: Using Blender and a specialized "Rain Generator" plugin, raindrops were modeled as a motion blur process. This approach simulates the aerodynamic oscillation and deformation of falling droplets.
- Background Collection: The team curated 327 aerial video sequences (~25,000 frames) using an SMRC-ct13 drone, ensuring a mix of urban, rural, and forest scenes captured from both horizontal and downward perspectives.
- Composition: A random sampling approach for rain masks was employed to ensure uniqueness. The final degraded images are a mathematical blend of clean backgrounds and simulated raindrop effects.
The UAV-Rain1k Dataset
The resulting UAV-Rain1k dataset consists of 1,000 high-resolution (1500 × 1000 pixels) paired images, with 800 allocated for training and 200 for testing. The dataset is characterized by its diversity, featuring four rain density labels (light, moderate, heavy, and irregular) and two primary camera perspectives (horizontal and downward). User studies indicate that these synthetic images achieve higher subjective realism compared to previous benchmarks like Raindrop and RainDS, which are tailored for ground-level imagery.
Experimental Evaluation and Results
The benchmark was used to evaluate six state-of-the-art deraining algorithms. Transformer-based models demonstrated superior performance over traditional CNNs, likely due to their ability to model non-local information necessary for reconstructing occluded regions.
- DRSformer: The top performer, achieving a PSNR of 24.93 and SSIM of 0.9155.
- Restormer: A close second with a PSNR of 24.78.
Beyond visual quality, the practical utility of the dataset was validated using YOLOv8 for vehicle detection. Applying the deraining process significantly improved detection precision in rainy aerial footage, demonstrating the benchmark's value for downstream computer vision tasks.
Conclusion and Future Work
UAV-Rain1k provides a critical resource for improving drone vision in inclement weather. However, challenges remain, such as over-smoothing of fine background details and the appearance of artifacts in high-resolution processing. Future research should explore end-to-end systems that jointly optimize low-level deraining and high-level object detection tasks.
The dataset and code are publicly available at https://github.com/cschenxiang/UAV-Rain1k.
Summary updated