Dust to Detail: Restoring Sand-dust Images with Frequency-Guided Attention and Multi-Scale Features

Images captured during sand-dust weather conditions suffer from severe color distortions, loss of details, and noise, leading to significantly reduced visibility. These degradations disrupt key vision tasks, such as scene surveillance and intelligent transport systems. Existing restoration methods struggle to handle color distortion, recover fine details, limiting their applicability in real-world scenarios. In this paper, we propose a novel encoder-decoder framework that integrates spatial and frequency-based processing to address these challenges. At its core, our Frequency Selection and Weighting (FSW) module adaptively enhances informative frequency components while sup-pressing degradation-induced noise. We introduce a Multi-Scale and Directional Filter Bank (MSDFB) to distinguish local fine details from global structural patterns, enabling the retrieval of object shapes and textures. Additionally, we incorporate a Multi-Head Self-Attention (MHSA) mech-anism on the extracted feature map from the FSW module to focus on the most salient features. Extensive experiments demonstrate that our approach outperforms state-of-the-art methods across both synthetic and real-world datasets, including cross-dataset validation, achieving superior perceptual quality and significantly enhancing object detection performance in the restored images with lesser parameters.