Research On ECT Image Reconstruction Algorithm Based On Deep Belief Network

Electrical Capacitance Tomography(ECT)is a process tomography technique that began to develop in the late 1980 s.It has the advantages of fast safety,non-invasive and low price.Sensitivity soft field is a key problem leading to low reconstruction quality in ECT technology.And the speed of image reconstruction can not meet the application in some fields.So the quality and speed of image reconstruction is the key problem of ECT technology in research.The Deep Belief Network(DBN)is a kind of network which is often used in the field of deep learning.It is widely used in nonlinear systems,contains many hidden layers,can extract features layer by layer,and has strong learning ability for complex functions.The DBN is applied to ECT image reconstruction to improve the quality and speed of image reconstruction by avoiding the solution of sensitivity matrix.This paper analyzes the basic principle and overall structure of the ECT system in detail,introduces the common algorithms applied in ECT,and compares the imaging quality.To improve the quality of image reconstruction,the BP-DBN algorithm is proposed to realize the nonlinear relationship between capacitance value and dielectric constant in the detected field by DBN deep nonlinear network structure.BP algorithm is used for reverse fine-tuning.the BP-DBN is also improved.the adaptive step size(AS)is introduced into the contrast divergence(CD)algorithm to solve the problem of finding global optimum with fixed step size and improve the image quality.The pseudo-newton method is used to speed up the convergence and reduce the training time in the fine-tuning stage.Since the above DBN uses BP network for reverse fine-tuning,it is easy to fall into the local minimum,resulting in poor imaging quality,too long time.So ELM-DBN put it.Extreme Learning Machine(ELM)the connection weights between the input layer and the implicit layer are randomly generated and determined.The connection weights between the implicit layer and the input layer are no longer adjusted by iterative calculation,but by solving the equations to adjust the weights,greatly reduce the training time and improve the prediction accuracy.Compared with the BP-DBN,the ELM-DBN adjustable parameters are less and the model prediction effect is good,especially in the aspect of sample training speed.