A CNN-based image reconstruction scheme for complex-valued multi-frequency ECT
Y. F. Xiong1, L, Y, Zhu2, M. M. Zhang2 and W. Lu1
1School of Electrical Engineering, Shanghai University of Electric Power
Shanghai, 200090, China
2Tsinghua Shenzhen International Graduate School
Shenzhen 518055, China
wuluee@shiep.edu.cn
Purpose/Aim
Complex-valued multi-frequency capacitance tomography (CVMF-ECT) is a novel image reconstruction technology which use surface potential information in the sensitive-field to evaluate the dielectrics distribution in cross-sectional areas. CVMF-ECT can recover permittivity and conductivity distributions from impedance or admittance measurements, which has high values of practical applications such as non-invasive fault detection and medical imaging. However, the traditional iterative algorithms used to reconstruct images have numerous iteration times and a slow imaging speed. An optimal solution of this inverse problem is necessary to obtain a better reconstruction effect for the patterns with complex dielectrics distribution.
Experimental/Modeling methods
In this paper, a convolutional neural network (CNN) based image reconstruction algorithm is proposed to reconstruct the permittivity and conductivity distributions of different combinations of gas/liquid dielectrics at multiple frequencies. COMSOL Multiphysics finite element software was used to obtain the surface potential data for the reconstructed images, as well as to extract a representative sample of typical dielectrics combinations, and to obtain the capacitance values between electrode pairs and the theoretical dielectrics distribution matrix. The sample data was then normalized to train the network model. To verify the feasibility of the algorithm on CVMF-ECT image reconstruction, the structural similarity index measurement (SSIM) was used.
Results/discussion
By comparing to other popular image reconstruction methods, e.g., Tikhonov regularization and iterative Tikhonov, using CNN-based image reconstruction scheme always has the highest SSIM value for typical dielectric combinations such as gas as background fully filled with liquid, gas as background fully filled with gas, liquid as background fully filled with liquid, and liquid as background fully filled with gas.
Conclusions
This paper provides an effective algorithm for solving the inverse problem in image reconstruction of CVMF-ECT. The testing results confirm that the proposed algorithms in this paper can draw the key features of the original image more accurately and quickly, such as size, location, and movement of the dielectrics.