Cascade deep polarization network for precise image semantic segmentation

Optical polarization imaging technology provides multi-dimensional light field information, encompassing spatial details and polarization data, which can be exploited for image semantic segmentation for target scene analysis. Most recent works focus on the development of neural networks with separate simple preprocessing steps to deal with the raw polarization images, which limit the accuracy of semantic segmentation. This paper proposes a novel, to the best of our knowledge, method, dubbed cascade deep polarization network (CDPN), to improve the performance of semantic segmentation by integrating preprocessing modules directly into the end-to-end deep learning work. The raw input data include the angle of linear polarization, degree of linear polarization, and a set of Stokes parameters. The multi-dimensional feature maps are extracted from the raw data through the image denoising, fusion, and enhancement modules, which are then concatenated with a backbone network to obtain the segmentation results. By collaboratively training the preprocessing modules and backbone network with self-supervised loss functions, we strive to find out the optimal segmentation solution. Experimental results show that the proposed method can effectively improve the segmentation accuracy, while maintaining fast computation speed.