Imaging Through Dynamic Thick Scattering Medium Based On Deep Learning

Optical imaging through scattering media has attracted considerable attention due to its important research value in biomedical and remote sensing mapping fields.The traditional scattering media imaging methods are limited by the optical memory effect and the high requirements of the complicated optical system,which makes it difficult to solve the imaging problem caused by dynamic thick scattering media.In recent years,the technology of scattering imaging based on deep learning provides the possibility to achieve imaging through dynamic thick scattering media in complex environments,due to its advantages of no measurement transmission mode,fast imaging speed,and no field of view limitation.In this thesis,the transmission model of dynamic thick scattering medium was used to calculate the speckle image information under different parameters.Based on the deep learning technology,a neural network model based on the Ghost Module was designed.The simulation and experimental studies of imaging through dynamic scattering media with different concentrations and thicknesses were carried out,and the restoration effects of speckle images under three neural network models were analyzed.The main research contents and results of this thesis are as follows:1.Based on the multi-phase screen model,combined with the absorption and scattering processes and the spatial and temporal transformation characteristics of the dynamic scattering media,the absorption factor and rotation factor were added to the original phase screen to establish a dynamic scattering media model.the light propagation in the dynamic scattering media was simulated and the corresponding speckle image was collected for the neural network training.2.Combined with the advantages of the Ghost module that can generate highly efficient redundant information,a neural network model for dynamic thick scattering medium imaging was designed based on the encoder-decoder structure.The imaging quality and imaging speed of the Ghost Module-based neural network model and the convolutional neural network model and the deep separable convolutional neural network model were compared and analyzed.The results showed that the reconstruction of the simulated speckle image of the neural network model proposed in this thesis was the best,and the time required to reconstruct 100 speckle images was 3.4%more than that of the deep separable convolutional neural network.However,the deep separable convolutional neural network with the fastest imaging speed performed the worst in terms of imaging quality,and the convolutional neural network with the slowest imaging speed and imaging quality was in between.3.An optical system was designed and built for speckle image acquisition in dynamic thick scattering media.The experimental optical system was capable of speckle images acquisition and real-time calibration of the changes of dynamic scattering media.The reconstruction of three neural networks in dynamic scattering media of different concentrations and different thicknesses was studied,and the imaging quality was compared by two evaluation indexes of Peak Signal to Noise Ratio and Structural Similarity Index Measure.The results showed that the reconstructed image quality of the neural network model proposed in this thesis was the best in the case of different concentrations and thicknesses of dynamic scattering media,the deep separable convolutional neural network model was the worst but the fastest,and the convolutional neural network model was in between but the slowest.Compared with the other neural network,the neural network model constructed in this thesis was more suitable for the imaging study of dynamic thick scattering media.In this thesis,a neural network model for imaging through dynamic thick scattering media was designed,and simulation and experimental research were carried out,which providing new ideas and methods for imaging through dynamic thick scattering media.