Multi-Patch De-raindrop Transformer for UAV images

Due to the complexity of aerial photography environments during rainy conditions, rain droplets randomly adhere to the lens, significantly degrading image quality. Therefore, this paper proposes a de-raindrop method for UAV aerial images based on a fusion network of multi-patch and frequency Transformer. The network architecture consists of a three-stage image restoration network that applies a multi-patch segmentation strategy to optimize image patches of different sizes and positions. The proposed method leverages the strengths of Transformer algorithms by introducing a Frequency Attention Transformer Block (FATB). This block incorporates a frequency attention mechanism that decouples high and low-frequency components within the self-attention layers. By simultaneously focusing on both local and global information in the image, FATB achieves high-quality image reconstruction. Furthermore, to enhance feature fusion during image restoration, we introduce an Adaptive Feature Enhancement Module (AFEM). This module improves the representation capability of features across different stages, thereby further boosting the quality of image restoration. Experimental results demonstrate that the proposed method surpasses the state-of-the-art algorithms in raindrop removal, achieving an improvement of 0.41dB over the best existing algorithm while maintaining higher efficiency. Additionally, the method exhibits strong performance across other public benchmark raindrop removal datasets, indicating its broad applicability. In summary, this research not only advances the field of UAV image de-raindrop but also provides clearer and more reliable images for subsequent visual tasks.