Subspace-guided GAN for realistic single-image dehazing scenarios

Single-image haze removal is an essential preprocessing phase in many object detection and segmentation approaches. Recently, end-to-end deep learning-based approaches have dominated the field of single-image dehazing because of their superiority in recovering clear images corrupted by different types of degradation. However, training an effective dehazing network remains challenging, particularly in the absence of high-quality realistic training datasets. In this paper, a novel approach called a subspace-based dehazing generative adversarial network (SuDGAN) is proposed. Traditional training methods attempt to apply changes to pixel intensities, whereas SuDGAN adopts a novel training approach using existing synthetic datasets to learn the adjustment of subspace components related to haze. This approach enables the network to learn more discriminative haze-aware features and focus on adjusting the components that are more affected by haze (luminance) while preserving those that are less influenced by haze (structure). The proposed SuDGAN, along with several state-of-the-art approaches, is evaluated on various challenging synthetic and realistic datasets using haze-related and traditional evaluation metrics. The experimental results demonstrate the efficiency of SuDGAN in removing haze and producing visually pleasing results. Furthermore, the results show that SuDGAN has clear quantitative and qualitative improvements over most state-of-the-art dehazing approaches.