Feature Fusion Dehazing Network Based on Parallel Attention
Leyu Ma, Weili Ge, Ruowei Fang · 2024
Images captured in haze-affected weather frequently exhibit many issues, including the degradation of image features, distortion of colour, and the loss of local detail. Previous deep learning based image dehazing models have shown unsatisfactory performance on real-world non-uniform haze datasets. Most existing models are unable to combine features of different scales, resulting in poor performance on real datasets. In this paper, we proposed a parallel attention-based feature fusion dehazing network. We designed a multi-scale feature fusion module and a parallel attention module to improve the ability to handle haze of different concentrations and thicknesses. A comparison of the dehazing experiments conducted on different models with synthetic haze images and real haze images reveals that our model is more effective in handling haze of varying densities and thicknesses while retaining greater detail in the images.