Domain adaptive dehaing based on physical properties

Abstract

Deep learning-based single image dehazing has advanced significantly, yet models trained on synthetic data struggle in real-world scenarios. Tо address this cross-domain gap, we propose a Synthetic-to-Real Dehazing framework comprising two key components: 1) A domain adaptation network that generates Synthetic-to-Real hazy images by learning real haze characteristics through depth-transmission map correlations, and 2) A physics-guided dehazing network based on the atmospheric scattering model. Crucially, our framework requires no real hazy data during dehazing training. Experiments demonstrate our framework's superior cross-domain dehazing generalization.