Scattering Polarization Target Reconstruction Based On Deep Learning

Optical imaging technology,as a visual way for people to understand the world,has become an important direction in the development of modern science and technology.With the rapid development of science and technology,how to solve the visual imaging of the scattering media in complex scenes and obtain effective information for target reconstruction has become a hot topic of current research.The scattering effect of light is unavoidable in the transmission process,and the target information carried by light will be distorted.Therefore,the traditional optical imaging technology is limited,resulting in imaging blur,speckle and other problems.In recent years,polarization information and the three basic properties of light intensity,phase and wavelength are applied more and more widely.Polarization information has been proven to be a powerful means of information transmission in scattering medium,and it plays an active role in remote sensing,underwater imaging,and biological tissue imaging.Polarization not only reflects the precise information of the target in the scene,but also reflects the characteristic distribution of the scattering medium.However,when the scattering intensity or transmission distance of the environment increases,the polarization imaging method based on traditional physics is difficult to obtain a clear target image.On the other hand,data-driven computational imaging has become a hot spot.Computational imaging breaks through the equipment and optical limits of traditional imaging systems to obtain target information.It realizes the transformation from physical model to data model and reconstruct complete object information.Therefore,this thesis mainly uses the advantages of both polarization information and deep learning algorithms,and combines them to solve the problem of scattering imaging in complex environments.The main research contents of the thesis are as follows:(1)The polarization information is used to reconstruct target information in actual real environment.We introduce the relevant skills of data learning and the basic theory of convolutional neural network,and design the neural network model for training.We have utilized the scattering imaging principle based on learning to reconstruct the target.The real haze environment was built based on Monte Carlo algorithm,and the linearly polarization light was emitted to obtain the component of the difference between horizontal and vertical polarization as the speckle learning sample,and then input to the MU-DL-Net network to train the model.Finally,the untrained polarization speckle was used to reconstruct the target.The peak signal-to-noise ratio and structural similarity index were used as the evaluation indexes of image reconstruction to observe the clarity of target reconstruction.We have also studied the generalization ability of the network model and the effect of speckle target reconstruction for objects with different polarization characteristics.(2)The comparison between intensity and polarization reconstruction target is studied and the effect of polarization model is tested by using different polarization information.Intensity speckle training model was received by the system incident unpolarized light.By comparing the reconstruction results of intensity information and polarization information,the effectiveness of polarization information is proved.We have also studied different polarized light incidences to get the testing speckles and input them into the polarization model to reconstruct the target.The results show that the polarization information has robustness.