In this paper, we present an efficient and robust physics-embedded deep learning framework for image reconstruction in electrical capacitance tomography (ECT), where conventional inversion algorithms are fundamentally challenged by the ill-posed and nonlinear nature of the capacitance-to-permittivity mapping and the soft-field effect that blurs dielectric interfaces. The proposed dual-driven framework integrates data fitting with physics-based constraints derived from the ECT sensitivity field to improve reconstruction consistency and robustness. A two-stage architecture is developed: first, Kolmogorov-Arnold Networks (KANs) establish a nonlinear mapping from capacitance measurements to an initial permittivity image; second, a lightweight refinement network with a local high-frequency attention mechanism enhances interfacial details and mitigates boundary blurring by exploiting correlations among neighboring pixels. To further enforce physical consistency, a multi-objective loss function incorporates sensitivity-field priors, including symmetry, distance-dependent attenuation, and boundary conditions. Simulation results demonstrate strong reconstruction performance, achieving an average correlation coefficient above 0.9423, a relative squared error of 0.0103, and a PSNR of 34.95 dB. The framework effectively preserves fine structural details while improving accuracy and stability. Although the current study is limited to simulation data, the present work highlights the significance of physics infusion in deep architectures for constructing reliable and potentially deployable ECT reconstruction models.
Li et al. (Sun,) studied this question.