Regionally Routed Attention U-Net With Latent Fading Kalman Filtering for Real-Time EIT

Le résumé fourni par la source

Real-time electrical impedance tomography requires reconstruction algorithms that reconcile high structural fidelity with strict latency constraints. We propose RNU-Net, a high-performance, end-to-end time-difference EIT reconstruction framework based on a specialized 1-D ConvNeXt architecture augmented with regionally routed attention. To enable robust mapping, the model was trained on a large-scale corpus of 100 000 samples generated via EIDORS, encompassing inclusions of varying sizes and conductivities with canonical geometries (circles, squares, and right-angled triangles). To further enhance robustness in low-SNR environments, the framework integrates a latent fading Kalman filter for dynamic sensor-domain denoising. On the canonical synthetic test set, RNU-Net attains a high reconstruction accuracy (SSIM = 0.9738). Furthermore, we rigorously evaluate the model’s generalization capacity through a transfer learning strategy using a novel stochastic high-order Fourier benchmark, which introduces topological fusion to mimic complex structures. Evaluations on physical tank data and independent test set (KIT4) demonstrate that the framework maintains robust cross-domain adaptability. In terms of efficiency, the framework achieves an inference throughput of approximately 54.5 FPS on a mobile GPU and 12.4 FPS on a standard CPU, providing a practical solution for high-precision, real-time EIT monitoring.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.