Multistage Progressive Single-Image Dehazing Network With Feature Physics Model

Deep learning (DL) has recently shown superior performance for single-image dehazing. Most deep dehazing networks either estimate the parameters of the atmospheric scattering model (ASM) through a convolutional neural network (CNN) or predict the hazy-free image without physical prior. These typical approaches ignore the dehazing mechanism in feature space. In addition, single encoder–decoder architecture is widely utilized to construct dehazing networks, but it may suffer from limited semantic-level multiscale contextual dependencies. To tackle these issues, we first propose a novel physical model-based feature enhancement dehazing (PFED) block, which consists of a feature enhancement (FE) block and a feature dehazing block with two fully convolutional subnetworks. Such a fully convolutional module can improve the adaptivity of nonuniform feature dehazing. Based on the PFED block, we develop a multistage progressive dehazing network (MSPD-Net), which progressively removes haze in a multistage architecture. In addition, the selective kernel (SK) feature fusion scheme is used to carry out cross-scale and cross-stage fusion, which can enable the model to capture the intra- and interstage interactions, respectively. Extensive experiments on three popular datasets demonstrate that MSPD-Net is comparable or even superior to the state-of-the-art (SOTA) methods. Specifically, MSPD-Net exceeds the Transformer-based DehazeFormer-M 1.389 dB on the SOTS-indoor dataset. Furthermore, a series of ablation experiments prove that the key components of our method can boost performance effectively.