Deep learning-based polarization feature retrieval from a single Stokes vector

Abstract

Polarization is capable of probing microstructures and has unique sensitivity to fibrous anisotropic structure. Polarimetric imaging has demonstrated promising potential in diverse applications ranging from biomedicine, material science, and atmospheric remote sensing. The polarization properties of samples can be comprehensively described by a Mueller matrix (MM). However, the relationship between individual MM elements and properties of the sample is often not clear. There have been consistent efforts to derive polarization parameters from MM based on certain assumptions for better description of the samples, e.g., MM polar decomposition (MMPD), MM transformation (MMT) and MM differential decomposition. Usually, the MM imaging requires sequential measurements with different polarization states of incident light and the imaging process is time consuming. In addition, for movable samples, we cannot guarantee the consistency during the imaging. This may cause precision issues since the images cannot be well-registered. In this work, we built a statistical translation model to generate polarization parameters from a single Stokes vector which can be obtained by one-shot imaging. This will improve the imaging efficiency, simplify the optical system and avoid introducing errors by the image registration. In the model design, we adopted the generative adversarial network (GAN) where the generator is based on a U-net architecture. We demonstrated the effectiveness of our approach on liver tissue, blood smear and porous anodic alumina (PAA) film, and quantitatively evaluated the results by similarity assessment methods. The model can generate a parameter image within 0.1 second on a desktop computer, which shows the potential to achieve real-time performance.