CAT: A lightweight Color-aware Transformer for sandstorm image enhancement

Abstract

Sandstorm images are characterized by color casts and reduced contrast due to the presence of suspended sand particles, which significantly impacts the performance of high-level computer vision tasks. Recently, numerous deep learning-based methods have been proposed for sandstorm image enhancement. However, most of them are either ineffective or have excessive parameters. In this paper, we introduce a lightweight Color-aware Transformer (CAT) for sandstorm image enhancement. Specifically, we propose the Color Restoration Module (CRM), which integrates Channel Self-attention Transformer (CST) and Vision Transformer (ViT) to effectively correct color distortion by utilizing global channel information. Additionally, the Feature Refinement Module (FRM), composed of cascaded attention blocks, is designed to dynamically capture the most valuable information to refine the features. By leveraging the aforementioned modules, our network is capable of processing an image with 256 × 256 resolution in just 0.06 s, while maintaining a compact architecture of only 2.1M parameters. Through extensive experiments on both synthetic and real-world sandstorm datasets, we demonstrate that our CAT outperforms several state-of-the-art methods in terms of image quality and computational efficiency.