Multi-Scale Progressive Dehazing Network with Image Priors and Hybrid Attention Mechanism

Currently, most dehazing networks achieve good performance on synthetic images but perform unsatisfactorily when processing real hazy images; moreover, these methods usually ignore the application of prior knowledge about images. To address this issue, we propose a novel Multi-Scale Progressive Dehazing Network (MPDNet). We have found that the residual channel prior of hazy images contains rich structural information; thus, we design a Prior-Guided Block (PGB) and introduce it into the dehazing network. To better apply prior knowledge at different stages, we design an Attention-guided Feature Memory module (AFM) and a Multi-Scale Dehazing Unit (MDU), thereby achieving progressive image dehazing. AFM is used to explore the correlation between the current input and historical information, transferring features across different stages. MDU leverages the impact of dilated convolutions on the receptive field and combines a hybrid attention mechanism to extract more image information for restoring haze-free images. Extensive experimental results demonstrate that our MPDNet performs excellently in handling both homogeneous and non-homogeneous hazy images, comparable to existing state-of-the-art methods.