PDCISTA-Net: Model-Driven Deep Learning Reconstruction Network for Electrical Impedance Tomography-Based Tactile Sensing
Gang Ma,Haofeng Chen,Shuai Dong,Xiaojie Wang,Shiwu Zhang
DOI: https://doi.org/10.1109/tii.2024.3456443
IF: 12.3
2024-01-01
IEEE Transactions on Industrial Informatics
Abstract:Electrical impedance tomography (EIT)-based tactile sensor has shown great potential in human-machine interaction due to its low manufacturing cost, large-area scalability. However, challenges, such as limited spatial resolution, and artifacts in reconstructed images, hinder their effectiveness. In response, this study proposes a model-driven deep learning reconstruction network for EIT-based tactile sensing, named PDCISTA-Net. The framework integrates a preprocessing filtering module and a dual-channel iterative shrinkage-thresholding algorithm (ISTA). Unlike traditional ISTA, PDCISTA-Net employs a dual-channel structural network tailored to capture and represent block correlations and sparsity within impedance change distributions. This approach enables end-to-end training, where parameters, such as step size, nonlinear transforms, and shrinkage thresholds, are learned from generated training data. In addition, a novel filtering module based on the sensitivity matrix is introduced to enhance reconstruction quality by mitigating measurement noise. Numerical metrics and visual results show that PDCISTA-Net outperforms traditional Newton's one-step error reconstructor, total variation, ISTA-Net, and FISTA-Net methods with higher structural similarity index measure and peak signal-to-noise ratio. Ablation experiments verified the effectiveness of the dual-channel structure in improving reconstruction quality. Finally, we developed an EIT-based tactile system to validate the practical application of our approach. The results from real-contact detection demonstrate enhanced image quality and greater noise robustness compared to traditional reconstruction methods.