Image Reconstruction Based on Multilevel Densely Connected Network With Threshold for Electrical Capacitance Tomography

Abstract

Electrical capacitance tomography (ECT) is a real-time monitoring technology for the visualization of industrial dynamic processes. Due to the inherent nonlinearity and ill-posed nature of the ECT inverse problem, achieving fast and accurate image reconstruction remains a great challenge. A novel multilevel densely connected network with channelwise thresholds (MDCN-CW) is proposed, which adopts a deep-learning framework to implement a reconstruction process similar to iterative algorithms. MDCN-CW is a highly condensed framework that achieves efficient information transfer through dense connections between and within subnetworks, and soft thresholding is inserted into the subnetwork as a nonlinear transformation layer to eliminate unimportant features. Matching the purpose and structure of MDCN-CW, a phased strategy is used to train it. Each subnetwork is first trained with a stepwise individual strategy, and network parameters are fine-tuned after all subnetworks are integrated to improve the overall fit of the model. Simulation and experimental results show that, compared with existing deep-learning-based reconstruction methods and traditional algorithms, the proposed ECT image reconstruction network with soft thresholds has the advantages of simple structure, high imaging accuracy, and good generalization ability.