Image Reconstruction Algorithm Based On Deep Learning In Electromagnetic Tomography

Electromagnetic tomography(EMT)technology has the advantages of non-contact,vsdde range of applications and high safety.It9s widely used in mining,metallurgy,non-destructive testing and biomedieine.Its extensive industrial demand requires higher imaging accuracy.In the inverse problem of EMT,the solution of image reconstruction is often ill1posed and pathological.And the prior information is limited.How to improve the imaging accuracy and speed is a hot issue in this field.In recent years,artificial intelligence algorithms have been successfully applied in voice,image,translation and other fields.It guides us to explore the feasibility of using it to solve traditional EMT problem.We expect to learn imaging autonomously by learning algorithms and representative samples.And then it has a higher quality and speed of imaging.The main work content is as follows:(1)Inspired by the idea of dimensionality reduction in machine learning data preprocessing,the dimensionality reduction SVD algorithm is proposed.In this paper,simulation and experiment are done with three traditional algorithms and dimensionality reduction SVD algorithm.The advantages and disadvantages of this new algorithms are analyzed.(2)Based on the idea of sample training in deep learning,two deep learning image reconstruction algorithms SSAE+RBF and optimized fully connected(Optimized FC)are proposed to learn imaging in EMT.The details of two algorithms are described.The rationality of two kinds of network structures used to solve imaging problems is expounded theoretically.(3)Two types of 30000 samples are designed and simulated.It is used to learn imaging features and network parameters in the process of network training.A loss function is designed as the objective of network training optimization.The rationality of loss function is proved by theoretical deduction.(4)Through the contrast experiment of two network algorithms and two traditional algorithms on 2000 test samples,the imaging advantages of network algorithms are shown.By mixing different noise levels into the test sample set,the imaging contrast experiment is done.The anti-jamming ability of two network algorithms is demonstrated.And our deep learning algorithm has an advantage in computing speed with graphic processing unit.For the random 2000 test samples,which have similar type with training sample but doesn't learneed,both of two algorithms are superior to traditional algorithms in image reconstruction.It preliminarily verifies the feasibility of using deep learning algorithms to improve imaging speed and accuracy in EMT,In network imaging algorithm design,sample design and loss function design,it can be used for reference for further research in this field.