Spatial-frequency complementary fusion network for dehazing with multi-scale and attention modules

Single image dehazing is a challenging ill-posed problem. It aims to estimate the latent haze-free image from the observed hazy image. In recent years, learning-based methods have demonstrated their superiority in single image dehazing. However, most existing learning-based dehazing methods focus exclusively on spatial-domain features, largely overlooking frequency-domain information. To address this limitation, a novel end-to-end Spatial-Frequency Complementary fusion Network is proposed for single image dehazing. Its core idea is fusing complementary frequency-domain and spatial-domain information. To efficiently incorporate frequency-domain information, the network includes two meticulously designed modules: the Spatial-Frequency Multi-scale Module and the Spatial-Frequency Complementary Attention. The former achieves deep complementary fusion of spatial- and frequency-domain features through a branched architecture, strengthening feature representation and preserving image details. The latter modulates attention-enhanced spatial features in the frequency domain and employs an adaptive gating mechanism to emphasize informative regions, thereby enabling differentiated optimization of frequency-domain features and improving dehazing performance. Extensive experiments on synthetic and real-world datasets demonstrate that our method achieves competitive results, with notably better performance in color fidelity and detail retention.