MSAFFormer: A multi-scale attention fusion transformer for single image dehazing
Tongwang Zhang, Dongli Jia · Journal of Visual Communication and Image Representation · 2026
Single image dehazing remains challenging due to spatially varying degradation caused by atmospheric scattering. Existing methods often fail to model spatially varying transmission and perform haze-aware feature modulator, leading to over-smoothing or inconsistent restoration in heavily degraded regions. In this work, we propose MSAFFormer, an end-to-end hybrid Transformer framework that integrates multi-scale attention fusion with physics-guided feature refinement. By combining convolutional inductive bias with global self-attention, the model effectively captures both local texture details and long-range dependencies. To further enhance haze-related representation learning, we introduce the Haze-Aware Feature Modulator (HAFM), which adaptively reweights spatial-channel features using a transmission-aware gating mechanism and high-frequency enhancement. In addition, a Physics-Guided Feature Correction (PGFC) module embeds a differentiable atmospheric scattering prior into the deep feature space, enabling directional correction in heavily degraded regions while maintaining physical consistency. Extensive experiments demonstrate that the proposed method outperforms state-of-the-art approaches on multiple dehazing benchmarks.