Self‑Augmented Unpaired Image Dehazing via Density and Depth Decomposition (D4)

Image dehazing is essential for improving visibility in hazy conditions, but supervised methods that rely on synthetic paired data often overfit and fail to generalize to real‑world scenes. Recent unpaired approaches based on CycleGAN ignore the physical relationship between fog density, scene depth, and visibility.

To address these issues, the authors propose D4 (Dehazing via Decomposing transmission map into Density and Depth), a self‑augmented framework that separates the transmission map into a scattering coefficient (fog density) and a depth map. By estimating scene depth, the method can re‑render foggy images with varying thickness, providing rich data for self‑augmentation while training only on unpaired foggy and clear images.

The training pipeline consists of two symmetric branches: a dehaze‑to‑re‑haze branch and a re‑haze‑to‑dehaze branch. Losses include cycle‑consistency, adversarial, pseudo‑scattering‑factor supervision (randomly sampled scattering coefficients), and pseudo‑depth supervision, ensuring both content preservation and accurate physical parameter estimation.

Extensive experiments on benchmarks such as SOTS‑indoor, SOTS‑outdoor, and IHAZE demonstrate that D4 outperforms state‑of‑the‑art unpaired dehazing methods and even many supervised approaches, while using fewer model parameters and FLOPs. Qualitative results show more natural colors and less distortion compared with competing methods.

Beyond dehazing, D4 can generate foggy images with controllable thickness for image/video editing and can predict relative depth from clear images, although depth accuracy is limited compared with dedicated supervised depth networks. The method may over‑estimate transmission in extremely bright regions and can become unstable with low‑quality training data.

The authors conclude that incorporating physical‑model‑aware decomposition into deep networks improves robustness and can be extended to other low‑level vision tasks such as low‑light enhancement.

For more details, see the CVPR 2022 paper here and the released code at https://github.com/YaN9-Y/D4 .