FSDG-Net: A frequency-spatial parallel network with density guidance for image dehazing
Wenlu Yang, Hongyuan Jing, Yi Ren, Jinjin Hu, Mengfei Han, Mengmeng Zhang · Signal Processing Image Communication · 2026
Image dehazing is one of the key research tasks in the field of image restoration. However, most dehazing methods suffer from detail degradation and structural shifts which significantly reduce the dehazing accuracy. Moreover, the non-uniformity of real-world haze and the inherent difficulty of depth estimation often limit the effectiveness of existing methods. To address these problems, we propose a Frequency-Spatial Parallel Processing Block (FSPP), an architecture-level design that combines multi-scale spectral processing, spatial feature enhancement, and adaptive cross-domain fusion. Furthermore, for the non-uniform distribution of haze, we design a Haze Density Estimation Block (DEB). This block can learn the spatial distribution characteristics of haze and adaptively adjust the weights to achieve differentiated restoration. Finally, we introduce a Multi-stage Feature Fusion Block (MF) to promote feature complementarity and information flow between FSPP outputs of different depths. Experiments show competitive results on paired daytime benchmarks, while the cross-dataset and nighttime evaluations reveal remaining challenges under more complex conditions. The source code and instructions for reproducing the experiments are publicly available at https://github.com/xiao666lu/FSDG .