Study On Reconstruction Of Speckle Pattern For The Scalar And Vector Optical Fields Passing Through A Scattering Medium Based On Deep Learning

When light passes through a scattering medium,a speckle pattern is formed due to the interact between the light waves and particles in the scattering medium.The presence of scattering seriously impacts the quality of optical imaging in corresponding fields such as biomedicine,underwater imaging,remote sensing and surveillance.Therefore,It has become an important frontier topic to study the scattering effect of light field in scattering medium so as to achieve high quality imaging through scattering medium through eliminating the scattering effect.Some scattering imaging methods based on physical models,such as wavefront shaping,transmission matrix,scattering correlation and optical phase conjugation,have been studied extensively,but they still have disadvantages such as high computation cost,poor portability and sensitivity to noise.As a hot technology developed rapidly in recent years,deep learning has shown good results in solving a number of optical problems and has shown strong performance in scattering imaging.Most studies using deep learning for scattering imaging focus on scalar light fields with uniformly distributed polarization states.Because the phase,amplitude and polarization characteristics of the light field will be changed when the light field passes through the scattering medium,the imaging and restoration of a vector light field with spatially nonuniform polarization distribution through the scattering medium is of great value in both fundamental physical interest and practical application.In this paper,the reconstructions of scattering patterns of the scalar and vector optical fields are investigated by deep learning techniques respectively.The main research elements of this paper are as follows:1.The reconstruction of the original target image from scalar light scattering patterns is investigated based on deep learning.An experimental setup was completed to collect the speckle patterns of scalar light,and the speckle patterns were collected and pre-processed to construct datasets,which was fed into the convolutional neural network UNet for training.The experimental results confirmed that UNet network could effectively reconstruct speckle patterns and realize scattering imaging.2.A deep learning network model Trans＿CNN containing dual encoders is proposed to improve the restoration effect of scattering speckle image.In order to address the convolutional neural network limitation of local receptive field and insufficient perception of global information,resulting in partial missing of details in reconstructed images,a deep learning network model Trans＿CNN containing dual encoders is proposed,which fuses the information from the Transformer encoder and the CNN encoder information,and passes the fused encoding information to the decoder to obtain the reconstruction results,enabling the network to learn the maximum global features and local features from the scattergram.Experimental results indicate that the Trans＿CNN network combining the Transformer encoder and the CNN encoder has faster convergence speed and higher imaging quality in the reconstruction of speckle pattern..3.The simultaneous restoration of the orthogonal polarization components of vector light passing through scattering medium from a speckle pattern based on deep learning.Previous studies on deep learning for scattering imaging have mainly focused on scalar light fields with uniform polarization state distributions.An innovative proposal is proposed to reconstruct the speckle pattern of vector light through scattering medium based on deep learning,so as to realize simultaneous recovery of two orthogonal polarization component phase patterns of incident vector light fields from one speckle pattern.In the experiment,different data sets are encoded into the hologram as the phase diagram of the polarization components in the x direction and the y direction(or left and right circular polarization)respectively.The vector light field is generated based on the 4f system,and the corresponding speckle patterns are collected by passing the vector light through the scattering medium.In addition,a single-input,dual-output neural network,P-Dense UNet,was constructed,the speckle patterns and the orthogonal polarization component phase diagrams are input into the network for training,so that the network learns the mapping relationship between the input speckle pattern and the two orthogonal polarization component phase diagrams.The experimental results show that the P-Dense UNet network can simultaneously reconstruct the orthogonal x and y-direction(or left and right-circular polarization)polarization component phase diagrams from the speckle images.