Abstract: Aiming at the problem of poor reconstructed image quality in electromagnetic tomography (EMT) for metal defect detection due to the inverse problem unsuitability and pathology, an EMT defect imaging method based on the improved ResNet-50 algorithm is proposed in this work. Firstly, by simulating and modeling the eight-coil EMT inspection system, and followed by applying an electromagnetic field to the object under test and using sensor arrays to obtain the information of electromagnetic field distribution around it, a training set is constructed and the raw voltage data are preprocessed. And then, the nonlinear mapping ability of the deep residual network is utilized to complete the learning of the training set, and the training effect is evaluated by the test set. The experiments show that the improved ResNet-50 algorithm reduces the root mean square error by 87.10%, 81.63%, 57.79%, and 19.11% respectively, and improves the structural similarity index by 88.87%, 71.82%,16.24%, and 4.54% respectively, compared to Tikhonov regularization algorithm, Landweber iteration algorithm, VGG-16 algorithm, and improved ResNet-18 algorithm, and at the same time, can accurately restore the defect location, shape and size. Overall, the improved method significantly improves the reconstruction accuracy, quality and efficiency, confirming its superiority in EMT defect imaging.

Key words: electromagnetic metrology, electromagnetic tomography, improved ResNet-50, defect imaging, image reconstruction