The Densely Connected Pyramid Dehazing Network (DCPDN) is a novel end-to-end single image dehazing method that jointly learns the transmission map, atmospheric light, and dehazed image. It directly embeds the atmospheric scattering model into the network to ensure physics-driven dehazing. The method uses a densely connected encoder-decoder structure with multi-level pyramid pooling to estimate the transmission map, and a new edge-preserving loss function to optimize the network. A joint discriminator based on a generative adversarial network (GAN) is introduced to ensure the estimated transmission map and dehazed image are realistic. The network is trained using a stage-wise learning approach to improve convergence. The DCPDN architecture includes four modules: transmission map estimation, atmospheric light estimation, dehazing via the atmospheric scattering model, and a joint discriminator. The method is evaluated on synthetic and real-world datasets, showing significant improvements over state-of-the-art methods. The proposed method effectively preserves edges and avoids halo artifacts, and achieves better dehazing results with more accurate transmission map estimation. The network is trained using a combination of edge-preserving loss, L2 loss, and joint discriminator loss. The DCPDN is able to jointly optimize the transmission map, atmospheric light, and dehazed image, leading to improved performance in dehazing tasks. The method is validated on multiple datasets and shows superior performance compared to existing approaches.The Densely Connected Pyramid Dehazing Network (DCPDN) is a novel end-to-end single image dehazing method that jointly learns the transmission map, atmospheric light, and dehazed image. It directly embeds the atmospheric scattering model into the network to ensure physics-driven dehazing. The method uses a densely connected encoder-decoder structure with multi-level pyramid pooling to estimate the transmission map, and a new edge-preserving loss function to optimize the network. A joint discriminator based on a generative adversarial network (GAN) is introduced to ensure the estimated transmission map and dehazed image are realistic. The network is trained using a stage-wise learning approach to improve convergence. The DCPDN architecture includes four modules: transmission map estimation, atmospheric light estimation, dehazing via the atmospheric scattering model, and a joint discriminator. The method is evaluated on synthetic and real-world datasets, showing significant improvements over state-of-the-art methods. The proposed method effectively preserves edges and avoids halo artifacts, and achieves better dehazing results with more accurate transmission map estimation. The network is trained using a combination of edge-preserving loss, L2 loss, and joint discriminator loss. The DCPDN is able to jointly optimize the transmission map, atmospheric light, and dehazed image, leading to improved performance in dehazing tasks. The method is validated on multiple datasets and shows superior performance compared to existing approaches.