Adaptive learning-driven polarimetric dehazing imaging in a dynamic turbid environment

Polarization dehazing imaging has been an emerging topic to improve image quality by suppressing the scattering effects in complex scenarios. Most of these works are focused on the restoration of images in homogeneous scattering environments. Dehazing imaging in dynamic inhomogeneous scattering scenes becomes a challenging topic due to the potential applications in practice. In this work, we propose an adaptive learning-driven polarimetric dehazing imaging network (ALPD-Net) to effectively reduce the inhomogeneous scattering effects of suspended particles in a dynamic turbid underwater environment. According to the physical dehazing model, two polarization-related parameters at each pixel across the whole scene are accurately estimated in the proposed network with superior convergence and restoration performance. Additionally, sequential multi-frame polarization images are fused to extract more effective feature information for the image recovery of the target object. Experimental results demonstrate the effectiveness and robustness of the proposed method with better indicators compared to other dehazing models.