Data Efficient Single Image Dehazing via Adversarial Auto-Augmentation and extended Atmospheric Scattering Model

Supervised learning-based image dehazing algorithms are sensitive to degradation and training distribution, making them ill-suited for out-of-domain non-uniform restoration. We propose an adversarial auto-augmentation approach to address this limitation without explicitly collecting paired training data. Specifically, we generate images with a broad distribution representative of multiple domains by varying the degradation and color profiles achieved by leveraging new augmentation techniques, including mean-variance transfer, physically accurate atmospheric scattering model, and localized degradation generation. These techniques effectively account for non-homogeneous degradations, enhancing the robustness of the underlying degradation model. Apart from utilizing these synthetic negative images to train the underlying network, these also provide diverse image representations for enabling more effective contrastive regularization. In addition to the training modifications, we propose a frequency-based feature fusion mechanism that prioritizes semantic and structural information from the decoder and encoder. Finally, we incorporate depth and color attenuation priors to ensure perceptually pleasing and physically accurate restoration quality. To evaluate the efficacy of the proposed mechanism, we perform comprehensive experiments and obtain state-of-the-art (SoTA) results while achieving high fidelity and improving the performance of perception-based algorithms without fine tuning.