A color polarization demosaicing network based on sampling fusion and high-frequency information perception

Abstract

The color polarization filter array (CPFA) camera is widely used in various polarization-based computer vision tasks due to its ability to simultaneously capture both color and polarization information. However, due to the limitations of the CPFA's superpixel structure and the entanglement of information within the color and polarization channels, the demosaicking problem becomes highly ill-posed. Most existing methods only rely on a single sampling approach and ignore high-frequency polarization information, which often lead to color reconstruction errors and edge artifacts, thus affecting the quality of degree of linear polarization and angle of polarization. In this paper, a Color-polarization demosaicing network based on sampling fusion and high-frequency information perception is proposed to decouple and reconstruct color and polarization information. During low-resolution image sampling, the advantages of traditional interpolation and deep learning methods are integrated through a self-guidance residual compensation interpolation module, which provides richer cues for subsequent refinement. In the polarization reconstruction stage, a Stokes activation sub-network is introduced to leverage the high-frequency signals encoded in the Stokes vectors, thereby enhancing edge and detail recovery in the polarization intensity domain. Furthermore, a color-polarization weighted loss function is designed to jointly optimize the network by complementing information across different dimensions. Experimental results demonstrate that the proposed method achieves state-of-the-art performance in reconstructing polarization parameters and visualization. The accurately reconstructed polarization parameters provide a solid and high-quality guidance for subsequent tasks.