A deep neural networks-based image reconstruction algorithm for a reduced sensor model in large-scale tomography system

Image reconstruction for soft-field tomography is a highly nonlinear and ill-posed inverse problem. Owing to the highly complicated nature of soft-field, the reconstructed images are always poor in quality. One of the factors that affect image quality is the number of sensors in a tomography system. It is commonly assumed that increasing the number of sensors in a tomography system will improve the ill-posed condition in image reconstruction and hence improve image quality. However, as the number of sensors increases, challenges such as more complicated and expensive hardware, slower data acquisition rates, longer image reconstruction times, and larger sensitivity matrices will arise, resulting in a greater ill-posed condition. Since deep learning (DL) is capable of expressing complex nonlinear functions, the majority of research efforts have been directed toward developing a robust DL-based inverse solver for image reconstruction. However, no study has been conducted to solve the inverse problem and improve the quality of the reconstructed image using a reduced sensor model for a large-scale tomography system. This paper proposed an image reconstruction algorithm based on Deep Neural Networks (DNN) to investigate its feasibility in solving the ill-posed inverse problem caused by the reduced sensor model for a large-scale tomography system. The proposed DNN model is based on a supervised, feed-forward, fully connected, backpropagation network. It comprises an input layer, three hidden layers and an output layer. Also, it was trained using large data samples obtained from COMSOL simulation. The relationship between the scattered electromagnetic field measurement and the corresponding true electromagnetic field distribution vector is determined. During the image reconstruction process, the untrained scattered electromagnetic field measurement samples are used as inputs to the trained DNN model, and the model output is an estimate of the electromagnetic field distribution. The results show that the proposed DNN can accurately describe the distribution of electromagnetic field and boundary shape of phantom compared to traditional algorithms (LBP, FBP, Noser and Tikhonov), regardless of the size and number of phantoms within the monitoring area. Hence, the proposed DNN is more robust and has a high degree of generalization.