Polarization image demosaicking based on homogeneity space

Abstract

The advent of the micro-polarizer array allows focal plane sensors to capture polarized light information, resulting in the acquisition of subsampled polarization intensity images. The process of reconstructing a comprehensive representation of polarization information is called polarization demosaicking. Current polarization demosaicking methods are susceptible to encountering polarization artifacts along complex edges due to their failure to account for the intrinsic correlation between polarization channels. We define the homogeneity space by employing the metric neighborhood model to capture the underlying correlation among individual polarization channels. Using the homogeneity space as the foundation, we propose a DI-Fusion method that effectively adaptively diminishes polarization artifacts in the initial demosaicking performance. Additionally, our homogeneity space can be extended in other advanced polarization demosaicking techniques, effectively enhancing the overall demosaicking performance. The experimental results demonstrate the high efficacy of our method in reducing polarization artifacts and enhancing the initial polarization demosaicking performance across two widely recognized benchmark datasets. Our method significantly improves objective metrics and visual performance compared to the initial demosaicking performance.