All-in-one UAV image restoration via prompt-guided wavelet state space model
Authors:
- 1 July 2025
- journal article
- Published by SPIE-Intl Soc Optical Eng in Journal of Applied Remote Sensing
- Vol. 19 (03) , 1-15
Abstract
Unmanned aerial vehicle (UAV) images are widely applied in intelligent transportation, power line inspection, and disaster monitoring. However, it can be easily degraded by weather and environmental factors during acquisition. Several single degradation restoration techniques based on convolutional neural networks, transformers, and diffusion models have been proposed. In fact, these models excel at handling known degradation types but struggle to adapt to new degradations without extensive retraining. Recently, all-in-one approaches have provided a unified framework to address this challenge. Due to the unpredictability of degradation in real-world scenarios, these methods need further development. In this paper, we introduce a UAV perspective degradation dataset named UAVIR-5D, which encompasses five types of degradations: blur, raindrops, haze, low light, and noise. In addition, we propose a prompt-guided wavelet state space model. Specifically, we introduce the feature prompt block into the model to implicitly provide contextual information, dynamically guiding the image restoration process. Second, we utilize the wavelet transform to separate noise from the image content. This allows the low-frequency Mamba block to fully leverage its long-range modeling capabilities, whereas a high-frequency optimization block is introduced to further restore image details. Extensive experiments have demonstrated the value of our dataset and the effectiveness of the proposed model.Keywords
Funding Information
- Scientific Research Platform Construction Project of the Education Department of Liaoning Province (LJ232510143007)
- Shenyang Science and Technology Project (23-503-6-18)
- Fundamental Research Funds for the Universities of Liaoning Province (LJ232410143060)