Denoising diffusion null-space model for polarized image demosaicing

Real-time multi-polarization imaging using division-of-focal-plane (DoFP) polarimeters supports many vision tasks, but raw images from polarization filter arrays (PFA) require demosaicing. Sparse polarization sampling causes severe artifacts that reduce reconstruction quality. To address this issue, we propose a novel polarized image demosaicing method for optimizing DoFP imaging systems, which is built upon the denoising diffusion null-space model (DDNM). First, we demonstrate the application of DDNM to the polarized image demosaicing problem and analyze the conditions under which DDNM can potentially recover the polarization information from DoFP images. Guided by the above theoretical analysis, we further design an optimized DoFP pattern. Experimental results show our method outperforms baseline approaches and achieves competitive reconstruction performance.