From local to global: a multi-group feature enhancement network for non-uniform and dense haze removal

In the past few years, significant research on single-image dehazing has developed rapidly. Despite this effort, it is still hard to remove the dense haze completely, particularly in complex real-world cases. The real-world haze is non-uniform and varied (light or dense). In the non-uniform case, the structure of the image can be destroyed. Besides, the procedure of dense haze removal usually leads to color distortion, detail loss and structure blurring, which increases the difficulty of image restoration. To solve these problems, we propose a multi-group feature enhancement network (MGFEN) based on a global and local context fusion pattern to remove haze progressively. Unlike previous methods, we develop a global feature fusion (GFF) module which takes a more global perspective to extract features and performs attention fusion with high-frequency features obtained from the Laplace pyramid to effectively preserve structure information of the image and remove artifacts caused by non-uniform haze. We also design a feature residual enhancement (FRE) module to improve image details and boost color fidelity by enhancing effective residuals group by group. The Experimental results of different datasets show that our MGFEN establishes the new state-of-the-art performance for real-world non-uniform and dense haze removal both in objective metrics and visual quality with better structure and color recovery ability.