A novel high-speed photopolarimeter is presented

in which a incident light is divided into multiple beams by a special metallic grating that can generate both reflective diffraction and transmission diffraction. The light fluxes of the four 1st order diffracted beams are linearly converted into four electrical signals by a photoelectric conversion circuit. A multilayer linear neural network model is set up whose inputs are the electrical signals

and outputs are the Stokes parameters of the incident light. The mapping relationship between the electrical signals and the Stokes parameters can be determined by training the neural network. After the electrical signals are measured

the unknown Stokes parameters of the incident light can be calculated via a trained neural network. The testing results show that the mean deviation of the measured and theoretical Stokes parameters is less than 2% at

λ

=632.8 nm. This instrument is compact

easy to install and characterized by fast response

high precision and damaging-free in working states.