Illumination Adaptive Enhancement Model for Non-Uniform Low-Light Polarized Image

Since polarization imaging provides additional polarization information, it has been widely utilized in biomedical imaging, scratch detection, material characterization, and so on. However, polarized images in low-light environments suffer from non-uniform illumination, and the degree of polarization (DoP) and angle of polarization (AoP) are severely degraded. To address these issues, an unsupervised guided network for non-uniform illumination in low-light environments is proposed with two modules, an illumination-guided module, and a context-guided attention module, to adaptively capture luminance feature in non-uniform illumination conditions. The illumination-guided module accepts the original complex scene image and an illumination map as input to extract illumination features. The context-guided attention module is proposed at the jump connection between the encoder and decoder to assist the model in learning the original illumination features of the image. Furthermore, polarization loss is incorporated into network training to ensure the consistency of the polarization components and to restore the polarization information. In addition, we established a new unpaired low-light polarized image dataset for non-uniform illumination conditions. The experiments proved that this method enables adaptive enhancement of brightness of images and recovers the polarization information effectively.