Abstract
<title>Abstract</title> <p>Steel plate defect inspection based on electromagnetic tomography (EMT) has attracted increasing attention because of its ability to perform non-invasive imaging without direct contact with the inspected object. Nevertheless, accurate image reconstruction remains difficult owing to the highly ill-conditioned inverse mapping and the diffuse nature of electromagnetic fields, which often result in distorted defect boundaries and unstable reconstructions. Although deep neural networks have demonstrated considerable potential in EMT imaging, their predictions may deviate from underlying physical principles when trained solely from data. To improve both reconstruction accuracy and physical reliability, a physics-guided image reconstruction framework named PI-TransUNet is developed. The framework integrates a TransUNet-based inverse solver with a pretrained differentiable forward surrogate model, forming a closed-loop learning architecture. Spatial positional information is incorporated into the measurement sequence to strengthen topological awareness, while a physics-consistency objective is introduced to regularize the reconstruction process. Furthermore, a progressive weighting strategy is employed to stabilize optimization and facilitate the incorporation of physical constraints. Numerical simulations and laboratory experiments conducted on a nine-coil planar EMT platform demonstrate that the proposed approach achieves superior defect localization, enhanced boundary preservation, and improved robustness against measurement noise. Quantitative results indicate consistently high structural similarity and low reconstruction error, highlighting the effectiveness of PI-TransUNet for EMT-based steel plate defect imaging.</p>