Lightweight and Energy-Efficient Object Localization in Electrical Impedance Tomography via Task-Oriented Inference

We propose a lightweight, task-oriented inference framework for object localization in electrical impedance tomography (EIT) without solving the inverse reconstruction problem. Training data were generated by finite element method (FEM) simulations for Opposite and Adjacent current injection configurations. A simple feedforward neural network independently estimated radial and angular object positions as probability distributions. These distributions were visualized and quantitatively evaluated using the Wasserstein distance. Results show that the Opposite configuration produces more lo-calized, unimodal probability distributions than the Adjacent configuration by alleviating informational ambiguity in the central region. Experimental EIT measurements con-firmed robustness, and the proposed approach significantly reduced energy consumption from 3.09 Wh to 0.96 Wh, demonstrating over three times higher efficiency than con-ventional methods. This task-oriented framework enables reliable and energy-efficient localization for low-power EIT applications.