Two-stage real-world image dehazing method using physics-based dehazing network and contrastive learning generative adversarial network

Real-world image dehazing remains a challenging task due to the ill-posed nature of haze formation and the significant domain gap between synthetic and real foggy scenes. In this paper, a novel two-stage framework is proposed, integrating a Physics-Based Dehazing Network (PBDNet) with a Contrastive Learning-based Generative Adversarial Network (CLGAN). In the first stage, PBDNet is trained on synthetic hazy-clean pairs using the atmospheric scattering model, extracting interpretable and transferable physical priors. In the second stage, CLGAN leverages these priors to guide unpaired image translation between real hazy and clean images. The integration of contrastive learning further enhances the alignment of fog-invariant representations, improving dehazing stability and generalization. Extensive experiments demonstrate the effectiveness of our approach. On the SOTS-outdoor dataset, our method achieves a PSNR of 34.13 dB and SSIM of 0.9863, surpassing state-of-the-art methods. On the real-world RTTS dataset, it achieves a BRISQUE score of 17.54, indicating superior perceptual quality. Additional evaluations using FADE metrics and object detection tasks confirm the practical value of our method in real-world scenarios. These results validate the effectiveness of combining physics-based priors with contrastive learning for robust real-world dehazing.