Research On ECT Image Reconstruction Algorithm Based On Deep Learning

Electrical Capacitance Tomography(ECT)is a promising process imaging technology.The technology has the advantages of no radiation,non-intrusion,fast response,simple structure and low cost.In recent years,it has been widely used in the fields of flow pattern recognition,medical imaging and industrial imaging.ECT image reconstruction is a nonlinear ill-conditioned inverse problem.The traditional non-intelligent image reconstruction algorithm mainly linearizes the nonlinear problem or adopts the local optimization nonlinear processing method.Both processing methods will make the image distortion more.Therefore,based on the deep learning idea,this paper chooses the convolutional neural network for ECT image reconstruction,and proposes improved algorithms for its inadequacies.1.Aiming at the problem of low imaging accuracy caused by the large number of approximations of Alternating Direction Method of Multipilers(ADMM)in ECT image reconstruction and the difficulty of choosing network topology structure in ECT image reconstruction by deep learning,an algorithm combining ADMM with deep learning is proposed.Using the powerful learning ability of deep learning and choosing the optimal sparse basis and other algorithm family parameters in the process of network training,the problem that parameters are difficult to determine in ECT image reconstruction of ADMM algorithm family is solved,and the problem that network structure is difficult to select is solved by mapping ADMM algorithm family directly to deep convolution neural network.The experimental results show that the algorithm combining of ADMM model family and deep learning has higher image reconstruction accuracy.2.An improved convolution neural network algorithm based on adaptive regularization coefficient is proposed to solve the problem that the weight of Relu activation function can not be updated in the negative part due to the disappearance of gradient and the convolution neural network is prone to over-fitting.By introducing an improved SLU activation function,the problem of gradient disappearance in the negative part is avoided.On this basis,a loss function with adaptive regularization coefficient ? is added to solve the over-fitting problem of convolutional neural networks.The improved convolution neural network isapplied to ECT image reconstruction.The experimental results show that the generalization ability of the improved convolution neural network is enhanced and the imaging accuracy is improved.