Image Reconstruction for Electrostatic Tomography With One-Dimensional Prior Knowledge Based Residual Network

Electrostatic tomography (EST) can measure the flow parameters of the gas–solid two-phase flow based on the reconstructed charge image. The traditional image reconstruction methods cannot achieve high image quality because a lot of specific information is lost during the linearization process. Deep learning-based reconstruction methods can avoid the calculation of the linearization process. However, the prior knowledge of the relationship between boundary measurement data and the true image is rarely applied in the deep learning-based methods for EST. To solve this problem and improve the quality of reconstructed images, a novel 1-D prior knowledge-based residual network (PK-ResNet) is proposed. The whole network is divided into three parts, namely, the initial feature extraction module, the residual feature extraction module, and the image reconstruction module. The prior knowledge of EST is encapsulated as a loss function to enhance the performance of the network. The network is trained by 35 000 generated samples. Simulation and experimental results show that the correlation coefficients of the reconstructed images of PK-ResNet are larger than those of the traditional algorithms and some existing image reconstruction algorithms based on deep learning.