Electrical Impedance Tomography Image Reconstruction With Attention-Based Deep Convolutional Neural Network
Zichen Wang,Xinyu Zhang,Rong Fu,Di Wang,Xiaoyan Chen,Huaxiang Wang
DOI: https://doi.org/10.1109/tim.2023.3265108
IF: 5.6
2023-04-21
IEEE Transactions on Instrumentation and Measurement
Abstract:Electrical impedance tomography (EIT) is a promising functional and structural imaging method in process tomography. However, due to the "soft-field" nature and the high dependence on the prior information, it often suffers serious artifacts in quantitative analysis. Most recently, EIT image quality has significantly improved because of the state-of-the-art (SOTA) deep learning-based models in the aspect of solving the inverse problem, especially fully convolutional networks (FCNs) and V-Net variants. Despite their success, these deep convolutional neural networks (CNNs) have two limitations: 1) the long-range information transition is frequently lost and the reverse gradient often disappears in deep CNNs and 2) some novel skip connections, such as residual and dense connections, often occupy substantial computational resources. To overcome these two limitations, we propose -Net, a new neural architecture based on redesigned feature transited connections by the terms of: 1) a prereconstructor based on the iterative Newton–Raphson (NR) method, which maps the nonlinear function between the measurements and the initial images; 2) dual cascaded V-Nets are combined, which play the role of an encoder and a decoder, respectively; 3) a new parallel attention mechanism via channel attention and coordinate attention to learn the conductivity distributions and boundary-shaped feature separately; and 4) the lightweight skip connections reduce the computational resources (or accelerate the inference speed) of EIT imaging. The -Net is evaluated by using the multiphase flow industrial applications, and the results demonstrate that: 1) -Net has a better performance in shape reconstruction with sharp "corner"; 2) -Net could reconstruct the model accurately where it has some low-contrast conductivity distributions; 3) -Net enhances the quality of interfaces with the stratified flow; and 4) the pruned -Net achieves significant speedup compared with the VDD-Net or -Net. The analyses show that the average relative error is 0.05, the average correlation coefficient (CC) is 0.92, and the average structural similarity (SSIM) is 0.92 on the testing datasets. In addition, the average relative cover ratio is 0.97, and the average relative contrast (RCo) ratio is 0.98 on the testing datasets.
engineering, electrical & electronic,instruments & instrumentation