EPGV-UNet: A measurement-domain feature-guided network for electrical resistance tomography image reconstruction

Abstract Electrical Resistance Tomography (ERT) image reconstruction is a typical ill-posed inverse problem, in which blurred boundaries, structural distortion, and local artifacts can arise under complex target distributions and high-resolution reconstruction, thereby limiting the accurate recovery of target structures and boundaries. To address this issue, this study proposes an ERT image reconstruction network based on measurement-domain feature guidance, termed EPGV-UNet. The proposed network extracts measurement-domain features from raw boundary voltage measurements and transforms them into spatial attention information to adaptively modulate image-domain decoder features, thereby providing measurement-domain guidance for image-domain feature recovery. Furthermore, a modified EfficientNet-B5 encoder and a progressive reconstruction decoder incorporating gated attention skip connections and PixelShuffle feature rearrangement are employed to improve high-resolution feature recovery. Through multi-scale feature fusion and progressive spatial-resolution recovery, the proposed network enhances the representation of target structures, contours, and boundaries under complex target distributions. Simulation and experimental results demonstrate that the proposed method achieves favorable performance in terms of overall reconstruction quality, target-region recovery, boundary preservation, noise robustness, and distribution generalization. Moreover, the proposed method can effectively recover the position, shape, and boundary information of bubbles under practical measurement conditions, providing an image reconstruction approach for online visualization measurement of complex gas-liquid two-phase flows using ERT.