Color polarization demosaicking by a convolutional neural network

Opt Lett. 2021 Sep 1;46(17):4338-4341. doi: 10.1364/OL.431919.

Abstract

To address color polarization demosaicking problems in polarization imaging with a color polarization camera, we propose a color polarization demosaicking convolutional neural network (CPDCNN), which has a two-branch structure to ensure the fidelity of polarization signatures and enhance image resolution. To train the network, we built a unique dual-camera system and captured a pairwise color polarization image dataset. Experimental results show that CPDCNN outperformances other methods by a large margin in contrast and resolution.