Large-range polarization scattering imaging with an unsupervised multi-task dynamic-modulated framework

The adaptability of most polarization scattering imaging, such as supervised learning-related and physics-based algorithms, is still limited, especially in dynamic or variable scattering environments. More importantly, it is a challenge to obtain a sufficient amount of polarization data for practical applications. Hence, we propose a multi-task dynamic-modulated polarization de-scattering framework (MDPF) based on the unsupervised training strategy, which achieves real-time de-scattering for a single image, by using the physical transmission model of light in scattering media as well as gradient descent optimization. Our proposed framework can significantly improve adaptability to large-range scattering conditions, without large-scale datasets for training. In addition, the experiments in real haze scenarios verify the effectiveness and superiority of our proposed method, demonstrating the potential of the proposed method in real-world environments, such as remote sensing or underwater imaging.