Physical Information Aware Fusion Network for Electrical Resistance Tomography

Electrical resistance tomography (ERT) is a noninvasive, radiation-free imaging technique widely used in industrial, biomedical, and tactile sensing applications. However, the inherent ill-posedness and nonlinearity of the ERT inverse problem often lead to low spatial resolution and noise artifacts in existing methods. To address these challenges, we propose a physical information-aware fusion network based on a cascade learning scheme (Phy-FNET-C). Our approach employs a streamlined encoder-decoder architecture that achieves superior performance without the computational overhead of overly complex designs. Specifically, a hybrid dual-branch structure is designed to extract augmented and crucial features from both initial reconstructions and upsampled electrical resistance map. A physical information-aware fusion module is introduced to integrate detailed physical guidance from voltage measurements with spatial priors. Furthermore, an uncertainty-aware multitask learning scheme is developed to optimize global and edge information simultaneously. Finally, a phased cascade learning strategy further refines image boundaries. Across extensive simulations and physical experiments, Phy-FNET-C consistently yields clearer reconstructions and higher accuracy compared to state-of-the-art approaches.