Explainable and Reliable Image Dehazing: A Survey of Deep Learning Approaches

Image dehazing has witnessed significant advancements with the adoption of deep learning techniques, particularly convolutional neural networks, generative adversarial networks, and transformer-based architectures. However, despite improvements in visual quality and quantitative performance, existing approaches often lack reliability and interpretability, which limits their deployment in real-world safety-critical applications. This survey presents a comprehensive review of deep learning-based image dehazing methods with a specific focus on explainability and reliability. We systematically categorize existing approaches into prior-based, CNN-based, GAN-based, and transformer-based methods, and analyze their strengths and limitations. Furthermore, we highlight the emerging need for confidence-aware modeling and explainable frameworks to enhance trustworthiness. Our analysis reveals that while GANs and transformers improve perceptual quality, they remain largely black box in nature. The survey also identifies key research gaps, including the lack of standardized evaluation for interpretability and limited real-world generalization. Potential applications include autonomous driving, surveillance, remote sensing, and medical imaging. This work aims to guide future research toward developing robust, interpretable, and deployment-ready dehazing solutions.