Deep Image Refinement Method by Hybrid Training With Images of Varied Quality in Electrical Capacitance Tomography

Abstract

Electrical capacitance tomography (ECT) has been applied in many fields for process monitoring. Knowing the permittivity values of target dielectric objects is essential in lots of application scenarios. Conventional methods have difficulty in predicting permittivity values of dielectrics and cannot solve the dilemmas between image quality and time efficiency. A deep learning-based method is proposed to refine the images reconstructed by conventional methods, with a multi-level fusion layer added to derive the permittivity values of multiple (>2) different dielectrics in the sensing field. A hybrid training strategy is implemented by taking images of varied qualities as inputs to the deep neural network during training, which are reconstructed by non-iterative and iterative methods such as Linear back projection (LBP) and Landweber iteration. With this strategy, low-quality input images reconstructed by LBP can be refined to the equivalent level of quality as those by iterative methods such as Landweber iteration, with the time consumption significantly reduced. Simulation and experimental results confirm the effectiveness of the proposed method by comparing with the conventional reconstruction methods as well as other deep learning-based methods without taking images reconstructed by conventional methods as network inputs or without using the hybrid training strategy.