Deep learning scheme PSPNet for electrical impedance tomography

The inverse problem of electrical impedance tomography (EIT) is often a highly nonlinear ill-posed problem of image reconstruction and the imaging feature prior information is limited. Since traditional iterative algorithms are not able to process the blurred EIT images, a deep learning algorithm is proposed in this study that can autonomously learning important imaging characteristics by some representative samples to improve the image reconstruction quality. First, a PSPNet network based on MobileNet are preliminary proposed to the EIT image post-processing, some simulation experiments based on the EIDORS platform were designed to illustrate the effectiveness and universality of this algorithms. Then we used the NOSER algorithm to reconstruct images of different shapes of targets, including circle, triangle, square, and any combination of two shape targets. For these resulting images, we used PSPNet for postprocessing. Furthermore, we designed a tank test with the EIT systems. Through detection of single target and two target positions in three different positions, the proposed deep learning PSPNet algorithm can improve the Size Error (SE) index by 9.02% on average, and reduced the Position Error(PE) index by 1.59% compared with the traditional NOSER algorithm. It can be concluded that the deep learning scheme PSPNet can effectively recognize the object positions and recover the sharp contour of targets very well.