FGN: A Fully Guided Network for Image Dehazing
Mingye Ju, Fuping Li, Siying Xie · IEEE Signal Processing Letters · 2024
Multi-scale fusion strategies have proven their efficient and effective performance in image dehazing tasks. However, inadequate feature fusion can lead to underutilizing local and global features. To this end, we propose a Fully Guided Network (FGN) for image dehazing. Specifically, we design a novel multi-scale aggregate attention (MAA), which aims to fully utilize early multi-scale features to guide the subsequent learning of the network. To prevent information redundancy, we develop an efficient multi-scale gated fusion module (MGFM) to control the information flow of different feature maps in the decoder stage. Based on MAA and MGFM, CNN-Transformer dual-branch block (CTDB) is constructed as the basic unit to achieve more refined image reconstruction. Extensive experiments on synthetic and real-world datasets demonstrate that FGN surpasses other state-of-the-art dehazing methods in terms of quantitative scores and recovery quality.