Double Branch Image Dehazing Algorithm Based on Probabilistic Uncertainty Modeling

Addressing the problem that deep learning methods fail to consider the uncertainty inherent in image dehazing results in real-world scenarios, this paper proposes PP-VAENet, a single-modal image dehazing method based on probabilistic uncertainty modeling. The PP-VAENet method innovatively decouples the dehazing process into two stages: distribution estimation and consensus decision-making: (1) constructing latent space distributions through Conditional Variational Autoencoder (CVAE), then combining with Probabilistic Adaptive Instance Normalization module (PAdaIN) to generate multiple reasonable candidate dehazed images; (2) finally selecting the optimal dehazed image from candidate results through Maximum Likelihood Estimation. To enhance the network's feature representation capability when constructing distribution estimation, this paper designs the NIN-Mamba module to fuse local feature extraction with global dependency modeling. Comparative experimental results on public datasets demonstrate that PP-VAENet surpasses existing advanced single-modal image dehazing algorithms in both quantitative and qualitative results.