Research On Information Transmission In Scattering Medium Based On Deep Learning

When light propagates in a scattering medium,it interacts with particles in the medium what cause scattering.After multiple scattering,the original image and polarization are scrambled,and they appear as a disordered optical speckle on the detection surface.Nowadays,the information transmission of light in the scattering medium has become one of the hottest areas of optical research.Deep learning,as a technology that has emerged in recent years,has shown good performance in solving some optical problems.For the problem that light passing through the scattering medium,this article proposes a deep learning method.Firstly,we propose a application of deep learning in imaging through scattering media,we use three scattering media what have different scattering intensities and collect training data by experimental light path.Then,we build a convolutional neural network to reconstruct the image of the targets hidden behind the scattering medium.In addition,we build a densely connected neural network to improve the effect of reconstruction when the scattering is strong,and we prove the generalization and portability of the neural network by reconstructing speckle images of special handwritten numbers.The densely connected neural network we constructed can achieve imaging outside the range of the optical memory effect,and the high degree of redundancy of the speckle image information is confirmed by compressing the original speckle image.Secondly,For the phenomenon that the polarization state of polarized light will be scrambled after passing through a multiple scattering medium,we propose a deep learning method to recognize the polarization state of incident light with high resolution.We build a convolutional neural network to extract features reflecting polarization information from speckle which formed by polarized light passing through a scattering medium.In addition,we use polarized light with different initial phases to verify the robustness of the neural network to recognize polarization state.As the neural network can be trained by data up to an infinitely large volume in theory,this work exhibits a huge potential of generalized suitability for solving the problem of information transmission in light scattering.