A convolution neural network for reconstructing microgrid polarimeter imagery

Light vector polarization as a fundamental property of lightwave, can be used to effectively distinguish objects in complicated circumstance including surface shape and materials type and transmission medium. As shown, polarization imaging is an advanced information acquirement method which combines the light intensity image and light vector vibration behaviors, which is the direction of electric field of incident lightwaves. A typical microgrid polarimeter with a minimum repeat unit is composed of four pixelated linear polarizer demonstrating different vibration directions. Compared with full polarization information, the polarization image obtained has only one quarter polarization information in each direction. Thus, it will influence the accuracy of other information such as Stokes components, the degree of linear polarization (DoLP), and the angle of polarization (AoP). In this paper we propose a polarization demosaicing network to address the poloarized image demosaicing issue, which are then recovered into the original polarized image. This network aims to improve the accuracy of DoLP and AoP of the targets by adjusting three Stokes components of the network output. We already remove the batch normalization (BN) commonly used in CNN, and thus use a customized loss function to make it suitable for polarization image demosaicing. The experimental results show that network has demonstrated a best peak signal-to-noise ratio (PSNR) and then richer image detail and polarization target information than that of the original image.