Real-time polarization measurement based on spatially modulated polarimeter and deep learning

Abstract

Spatially modulated polarimeter (SMP) can be employed to obtain the polarization states of light by a single shot of intensity image, and it is stable, robust and easy to use. In order to simplify the calculation procedures and improve the measurement speed, we introduce the deep learning (DL) method into SMP. In this work, a convolutional neural network (CNN) based on VGGNet architecture has been trained by simulated samples that are generated by a theoretical model of a spatially modulated system. Then the trained CNN is utilized to evaluate the Stokes parameters of the input light beam according to its spatially modulated image obtained in the real experiment. The mean absolute error between the evaluated values and the true values of the normalized Stokes parameters can be less than 0.02, and the evaluation of each image takes only 0.03–0.04 s. Compared with the traditional Fourier analysis method, the DL method is close in accuracy while greatly enhanced in speed, making real-time polarization measurement achievable.