A Lightweight U-Shaped Transformer Network for Single Image Dehazing

Abstract

Single-image dehazing aims to restore a clear image from a single haze-degraded observation. Although recent deep learning methods have shown promising performance, they still struggle to balance long-range dependency modelling and computational efficiency. To address this issue, we propose a lightweight U-shaped Transformer network, termed UTNet. The proposed network adopts a U-shaped encoder-decoder architecture to progressively extract and reconstruct hierarchical features, enabling effective integration of shallow spatial details and deep contextual information. To reduce the computational burden of Transformer-based modelling, local-window spatial attention is introduced at each stage to efficiently model dependencies. Unlike Swin Transformer, UTNet leverages its hierarchical U-shaped architecture to enlarge the receptive field and progressively capture global contextual information. In addition, channel attention is incorporated in parallel with spatial attention to enhance feature representation further. Multi-scale auxiliary supervision is also introduced to improve optimization stability and promote consistent restoration from coarse structures to fine details. Moreover, gamma-based data augmentation is employed to improve the model generalization to real-world hazy images. Extensive experiments on the Haze4K and NH-Haze datasets demonstrate that the proposed UTNet achieves competitive dehazing performance while maintaining a compact architecture with fewer parameters and lower computational cost.