Hybrid model of ResNet and transformer for efficient image reconstruction of electromagnetic tomography

Abstract

Electromagnetic tomography (EMT) is an electrical tomography technique based on the principle of electromagnetic induction, aimed at studying the spatial distribution of materials with electromagnetic properties. However, traditional image reconstruction in EMT faces challenges such as ill-posedness and non-linearity, which lead to problems such as poor reconstruction accuracy, unclear positioning, and blurred image contours. To address these issues, this paper proposes a hybrid model that combines ResNet-18 and Vision Transformer (ViT) neural network to solve the inverse problem. In this hybrid model, the excellent feature extraction ability of CNNs and the global feature extraction ability of Transformers are fully utilized. By incorporating the last two residual blocks of ResNet-18 and the 8-head attention mechanism of ViT, the model effectively captures long-range dependencies and integrates features of different scales, enhancing robustness and generalization. The results after 250 iterations show that the hybrid model achieves better performance in EMT image reconstruction compared with the traditional algorithms. In addition, the robustness of the proposed model to noise and its reconstruction performance with random samples are tested, which confirms the reliability and generalization ability of the proposed model. Furthermore, an 8-coil EMT experimental system was built to verify the feasibility and effectiveness of the hybrid model and demonstrate its potential for application in EMT.