In this work, we propose a two-step color-polarization demosaicking network (TCPDNet). The network architecture of TCPDNet is inspired by
Morimatsu's interpolation-based CPFA demosaicking method. Our method consists of two well-studied sub-tasks: color demosaicking and polarization demosaicking. In each sub-task, demosaicking network parameters are shared across different kinds of input mosaicked data. We consider that even if the input mosaicked data differ in polarization orientations (for the color demosaicking task) or color channels (for the polarization demosaicking task), their inter-channel correlations should be the same. For example, the color demosaicking networks for 0° of Bayer CFA mosaicked data and 90° of Bayer CFA mosaicked data share the same parameters because their RGB correlations should be the same regardless of the polarization orientations. We further improve the performance of TCPDNet by introducing a reconstruction loss in the YCbCr color space. Using the loss function in the YCbCr color space, we expect the demosaicking networks to learn the inter-channel correlations effectively. Experimental results show that our TCPDNet outperforms existing methods by a large margin both quantitatively and qualitatively on
Tokyo Tech dataset and
CPDNet dataset.