A Dual-Channel Deep Learning Method for Demosaicing Based on an Optimized Color-Polarization Array

Polarization and color are two essential attributes of color-polarization imaging In recent years, the application of both polarization and color information has gained widespread attention across various image processing research domains. This paper proposes a novel color-polarization array design method by employing the non-dominated sorting genetic algorithm (NSGA-II) to generate mosaicked images with color and polarization information, followed by a dual-channel guided deep-learning reconstruction network to recover these attributes. This optimization method produces a compact 3×3 array with an innovative filter layout. This minimizes polarization angles and unit array size while maintaining the integrity of color-polarization information in the reconstructed images. In addition, the use of deep learning modules enables a more thorough restoration of missing information in mosaicked images from a new perspective. Experimental results demonstrate that our network can accurately demosaic the mosaicked output of the proposed array, effectively recovering both the color and the polarization data. Our compact 3×3 array outperforms the classical 4×4 polar-bayer array on benchmark color-polarization image datasets.Additionally, the reconstruction algorithm preserves the fidelity of color and polarization after demosaicing.