| Imaging through scattering media is of great significance in biomedicine,aerospace remote sensing,holographic imaging and other fields.Scattering imaging techniques in static scattering scenarios rely on stable optical systems and are limited by optical memory effects,making it difficult to achieve imaging recovery in dynamic scattering scenarios.In this thesis,a deep learning network model is proposed based on a computational holographic scattering imaging system.And based on this,scattering imaging research is conducted using speckles through a dynamic scattering media.The specific work is as follows.1.The propagation principle of light in random scattering media and the underlying principles of scattering imaging using deep learning techniques are investigated.Various neural network structures and their design ideas are explored,and then the similarity and time-varying characteristics of the speckle generated under dynamic scattering conditions are analyzed.2.An optical scattering imaging system and a scattering data acquisition system were designed and built to achieve automatic acquisition of scattering image data based on optical devices and MATLAB,while various types of scattering media and multiple original target image data sets enriched the scattering image data categories and provided sufficient training samples for the training of neural network models.3.A deep learning network model URSnet based on the “encoder-decoder” structure is designed and constructed.The scattering recovery performance of the model in static scattering scenes is investigated,and the effect of different loss functions on the scattering recovery is analyzed.The experimental results show that combining the advantages of Resnet and SEnet can improve the feature extraction and image reconstruction ability of the network model,while the loss function has a large impact on the convergence speed of the model and image reconstruction.4.A Generative Adversarial Network(GAN)was designed and constructed for scattering imaging recovery in dynamic scattering scenes.The recovery performance and generalization performance of the network model for binarized grayscale images are first explored.Then speckle data generated from detail-and contour-rich grayscale images of faces under different dynamic scattering media conditions are collected to study the recovery ability and generalization performance of the model for target images rich in grayscale information.The experimental results show that the GAN can reconstruct grayscale images of faces or simple images with sparse distribution by training with a large amount of speckle data,and it has high generalization and robustness to recover the reconstruction of speckle data in unknown dynamic scattering scenes. |
