Effective polarization-based image dehazing through 3D convolution network

The presence of numerous microscopic particles within haze contributes to the scattering of atmospheric light, leading to a notable degradation in the quality of captured images. Polarization, an intrinsic attribute of the light field, encapsulates a wealth of information concerning both targets and their surrounding environment and has demonstrated robust performance in dehazing real-world hazy images. In this study, we introduce an end-to-end dehazing network tailored for processing polarization-based hazy images. Leveraging the capabilities of 3D convolution, our novel POL-3D Encoder adeptly harnesses the correlation information situated within the polarization angle dimension, derived from fused polarization images. Furthermore, we integrate a Spatial Redundancy-Reducing module, effectively mitigating redundancy within feature maps across the spatial dimension. To assess the efficacy of our network and ensure its robustness, we curate a dataset rooted in the principles of polarization-based image formation. Experimental results conclusively demonstrate that our method achieves state-of-the-art performance on both synthetic data and real-world hazy images, thereby highlighting its potential for advancing image dehazing techniques.