Seismic Velocity Model Building And Optical Imaging Based On Deep Learning

In recent years,artificial intelligence and big data have swept across the globe,and technologies based on deep learning rise rapidly and have been widely applied to many fields.The inverse problem of mathematics and physics is an interdisciplinary subject,its main purpose is to obtain high-quality reconstructed images from observation through mathematical modeling.At present,intelligent imaging methods based on convolutional neural networks(CNNs)have become a hotspot of research in many fields.Aiming at the seismic velocity model building and optical imaging problems,in this thesis,we introduce deep CNNs.We use supervised learning algorithms,model-driven and data-driven combined optimization algorithms and unsupervised learning algorithms,respectively,to mine the potential priori information of the data to obtain high-signal to noise ratio imaging results.The details are as follows:First of all,to address the problems of underdetermined and high computational cost of velocity model building in seismic exploration,we propose a velocity modeling approach based on deep learning.A minimum function from multi-shot observation data to velocity model is established.The learning of samples enables the fully convolutional neural network to automatically mine the features and laws of the data,and fit the nonlinear mapping from the multi-shot observation data to the velocity models.The trained network does not require iterative optimization and can be directly used to estimate a new velocity model,which reduces the calculation time.Numerical experiments show that,compared with the traditional full-waveform inversion method,this method does not require an initial velocity model,and reduces human intervention.Then,in order to reduce the serious dependence of supervised learning methods on the training set,we propose a method for solving inverse problem based on the combination of CNN and physical model constraints.By learning from small samples,a CNN with the property of projector is obtained,which is combined with the projected gradient descent algorithm.The relaxation parameters are adopted to ensure the convergence of the algorithm.This method is applied to optical diffraction tomography.The numerical experiments show that,compared with the conventional total-variation regularization method and supervised learning algorithm,the proposed method is able to obtain highresolution reconstructions when the observation data is partially missing,and to further improve the imaging quality in combination with a more accurate nonlinear forward model.Finally,in view of the difficulty of obtaining the training set in practical applications and the continuity constraint of phase unwrapping,we propose a method via an unsupervised learning algorithm – deep image prior for phase unwrapping.Combining the advantages of model-driven and data-driven frameworks,the reconstructed phase images can be well represented by a neural network.Based on the consistency of physical observations,the mapping relationship from high-dimensional feature space to phase image space is learned by iteratively optimizing the objective function.The proposed approach does not require any training set and ensures that the reconstructed images satisfy the physical constraints.Numerical experiments show that compared with traditional model-based methods and supervised learning algorithms,this method can recover the phase images of organoids with a variety of complex structures and better maintain the smoothness of cell edges.