Physics-Guided Dual-Branch Network for Unsupervised Single-Image Dehazing

Abstract

Unsupervised dehazing with unpaired hazy and clear images lacks pixel-aligned supervision, making it difficult to simultaneously remove haze and preserve structural details. To address this issue, this paper proposes a physics-guided dual-branch network for unsupervised single-image dehazing. The Auxiliary Physical Reconstruction Branch refines a Dark Channel Prior (DCP) transmission estimate and reconstructs the hazy input through the atmospheric scattering model, thereby providing a physical constraint during training. The Main Image Dehazing Branch predicts the dehazed image. The Dynamic Prior Rectification Module (DPRM) compares features from the hazy input and the coarse DCP result and adaptively balances them according to their local consistency. Multiple Feature-Guided Large-Kernel Attention (FG-LKA) blocks enlarge the receptive field and capture contextual information. Experiments on SOTS, HSTS, I-HAZE and RTTS demonstrate favorable dehazing performance.