Spatial-temporal Prior Constrained Image Reconstruction Algorithm For EIT

Electrical Impedance Tomography(EIT)has advantages of high time resolution,non-invasion,and no radiation.It has broad application prospects in the fields of industrial and biomedical processes measurements.However,as a "soft field" reconstruction modality,reconstruction results from EIT always have low resolution and are susceptible to model errors and measurement noise.As a result,the image reconstruction algorithm is one of the hot topics in this field.Based on deep learning methods,this paper aims to the develop the space-time constrained image reconstruction algorithms for improving the quality of EIT on imaging stationary and non-stationary processes.The main contents as follows:(1)To relive the nonlinearity and ill-posedness of image reconstruction problem of EIT,a deep neural network is proposed following the statistical inverse and representation learning theory.The proposed network can produce probabilistic reconstruction results and is trained by a three-stage strategy.Numerical results show that the proposed network is significantly more accuracy and robust than the traditional image reconstruction methods of EIT,like Tikhonov,and can produces the uncertainty of imaging results.(2)To improve the accuracy of EIT on dynamic imagereconstruction,a spatialtemporal constrained dynamic image reconstruction network is proposed.The network is composed of a pre-reconstruction network and a spatial-temporal enhancement network.It can automatically learn the spatial-temporal prior information from the training dataset and improve the result quality.To improve the generalization ability of this model,a random data interpolation method is proposed,realizing the generation of large-scale dynamic datasets from small-scale static datasets.Experimental results show that the proposed network can achieve high-performance dynamic imaging,including filtering,smoothing and prediction,and the imaging accuracy is better than that of the Kalman filtering method.(3)To improve the speed and accuracy of EIT on imaging of non-stationary processes,a spatial-temporal constrained iterative neural network is proposed.Different from the previous dynamic imaging network,this network is based on the physical model of EIT imaging and adopts an iterative method to learn the temporal and spatial prior information.It has easy implementation,simple training,and high accuracy advantages.To verify the performance of the proposed network,a two-dimensional random data interpolation method is proposed.Based on static experimental data,the synthesis of complex dynamic process data is realized.The test results show that the speed and accuracy of the proposed iterative neural network are significantly higher than the traditional unsteady image reconstruction algorithm.