End-to-end physics-informed multi-branch GAN for enhanced DoFP polarization image reconstruction

Division-of-focal plane (DoFP) polarimeters deploying micro-polarizer array as the polarization state analyzer (PSA) possess the capacity to capture polarization properties of the scene target during a single snapshot and benefit from their rugged and compact designs. However, these systems acquire polarization measurements through spatial modulation, leading to inevitable spatial resolution loss and reduction in polarimetric accuracy. To overcome these challenges, we propose, to our knowledge, a novel approach by leveraging an end-to-end physics-informed residual generative adversarial network (GAN) for DoFP polarization image reconstruction. Our method enhances the reconstruction of intensity (I), degree of linear polarization (DoLP) and angle of polarization (AoP) directly from DoFP polarization images bypassing traditional interpolation methods that rely on interpolating intensity images from different polarization orientations. The network's architecture is tailored to simultaneously handle demosaicking and polarimetric reconstruction, thereby mitigating the inherent limitations of DoFP systems. Additionally, we utilize Grad-CAM for model interpretability, allowing us to visualize and understand the regions of the input images that the network focuses on during reconstruction. Experimental results demonstrate that our approach improves the quality of the reconstructed polarization images and enhances overall polarization accuracy.