Electrical engineering

Inverse problem

Pattern recognition (psychology)

Image resolution

Mathematics

Algorithm

Fuse (electrical)

Iterative reconstruction

Engineering

Residual

Mathematical analysis

Tikhonov regularization

Artificial intelligence

Computer science

Feature extraction

Computer vision

Convolutional neural network

IEEE Transactions on Instrumentation and Measurement

Electrical resistance tomography (ERT) is an effective visualization and analysis tool for multiphase flow process through an array of boundary electrodes. However, the existing deep learning network for the image reconstruction of ERT with the sparse information flow and gradient flow is not well trained, and the accuracy of reconstructed images cannot meet the increasing demands. In order to solve this problem and improve the accuracy of the reconstructed image, a new imaging algorithm that consists of initial imaging module, feature extraction module, image reconstruction module, and residual module is proposed based on dense connections. Four dense blocks with dense connections are adopted to fully fuse the features between feature extraction module and image reconstruction module, which mitigate the problems of information vanishing of the higher layer and gradient vanishing of the lower layer. From the perspective of information flow and gradient flow, why dense connections have a good performance of promoting a deep network training is explained, which provides a theoretical support for the application and development of dense connections. Compared with the Tikhonov regularization method, total variation method, and the networks without dense connections using extensive numerical simulation and experiment results, the imaging accuracy of the new proposed algorithm is improved. Besides, for the ERT cases considered here, the spatial resolution of V Dense Net (VD-Net) (3.12%-4%) is better than that of the traditional image reconstruction (10%-15%).

Electrical Resistance Tomography Image Reconstruction With Densely Connected Convolutional Neural Network