A Review of Real-Time Single Image Dehazing Architectures for Embedded Vision Under Adverse Weather Conditions

Single image dehazing plays a pivotal role in enhancing visual quality for computer vision systems operating in outdoor or adverse weather conditions. Applications such as autonomous driving, aerial surveillance, smart monitoring, and environmental analysis rely heavily on clear visibility. Haze, caused by atmospheric light scattering, degrades image contrast, sharpness, and color fidelity, significantly impacting the performance of downstream tasks. This review presents a comprehensive overview of single image dehazing techniques developed over the past decade, bridging prior-based physical-model-based methods with recent learning-based approaches. prior-based dehazing methods, such as the Dark Channel Prior (DCP), Color Attenuation Prior (CAP), and Retinex theory, are classified as domain-based techniques. These approaches leverage handcrafted assumptions to estimate scene transmission and atmospheric light, offering computational efficiency but often struggling with dense or non-uniform haze scenarios. In contrast, learning-based methods including convolutional encoder-decoder architectures, multi-scale feature fusion, generative adversarial networks (GANs), and emerging transformer-based model’s achieve superior performance by learning haze-relevant representations directly from data. These models generalize well across synthetic and real-world domains and support end-to-end optimization. We evaluate these methods on standard datasets, using metrics like PSNR and SSIM to assess pixel fidelity and perceptual quality. Furthermore, we analyze the impact of various loss functions on training dynamics and restoration outcomes. A particular emphasis is placed on real-time, hardware-friendly dehazing models for edge deployment, considering power, latency, and memory constraints. Finally, we highlight ongoing challenges including generalization to unseen haze conditions, color distortion, limited real-world data, and training complexity. Future directions include transformer efficiency, unsupervised and semi-supervised learning, multimodal fusion, lightweight network design, and domain adaptation. This review serves as a timely resource for advancing robust and deployable dehazing solutions.