International Society for Industrial Process Tomography

Deep Learning with Classical Image Reconstruction Algorithms for Electromagnetic Tomography

Pengfei Zhao, Ze Liu*, Jun Xiao, Jiwei Huo and Yong Li

School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing, China

*Email: zliu@bjtu.edu.cn

ABSTRACT

Deep learning electromagnetic tomography is a new imaging method for tomography, it generates data, constructs images according to electromagnetic tomography, builds networks and trains samples in the way of deep learning. Early research has found that deep learning electromagnetic tomography could learn features from tens of thousands of samples. For specific flow patterns (multi-objects and single connected area), the image correlation coefficient increased from 0.2 to 0.7 compared with the classical imaging algorithms. But the imaging performance is unsatisfying when the test set has low similarity with the samples already trained. It is not realistic to gain and train all the samples. This is actually a manifestation of the generalization ability of deep learning, referring to the adaptability of fresh samples. However, there are intrinsic factors in tomography to increase generalization ability. In electromagnetic tomography, classical imaging algorithms, such as linear back projection, Tikhonov regularization, Landweber iteration, etc., have universal imaging effects on most flow patterns. Calculation data of classical imaging algorithms are added to the training and testing of deep learning. Reconstructed images of the test set and generalization set show that deep learning integrated with classical image reconstruction algorithms still preserve the general information of original image compared with simply using either when it comes to fresh samples. This method also has a positive reference for both process tomography and deep learning area.

Keywords Electromagnetic tomography, Deep learning, Image reconstruction algorithms, Landweber iteration, Generalization ability

Industrial Application General