Transfer learning-driven inversion method for the imaging problem in electrical capacitance tomography

Low-quality tomograms constrain the potential of the electrical capacitance tomography technology. In order to break through this bottleneck and innovate reconstruction algorithms, the deep transfer learning prior (DTLP) is introduced in this study, which is coupled with the imaging physical mechanisms and the domain knowledge modeled by a new regularizer into a new imaging model. The proposed imaging model is solved by a new optimizer in a simpler and less computationally expensive way. A new deep transfer learning method is developed to infer DTLPs by synergizing deep convolutional neural network with extreme learning machine (ELM) based on the collected multi-fidelity training samples. The training of the ELM is formulated into a new bilevel optimization problem, and a new nested optimizer is proposed to solve the problem. The quantitative and qualitative evaluation results confirm that the new method shows performance advantages over popular imaging methods in terms of detail restoration, noise immunity, artifact removal and edge preservation. The proposed imaging method synergizes deep transfer learning with imaging physical mechanisms, providing new opportunities and insights for unlocking the potential of the measurement technique and achieving better reconstructions.

Automated summary

In this study, a deep transfer learning prior (DTLP) is introduced to overcome the limitations imposed by low-quality tomograms on electrical capacitance tomography technology. The proposed imaging model, which incorporates imaging physical mechanisms and a new regularizer, is solved in a simpler and less computationally expensive way. A new deep transfer learning method is developed and shows performance advantages over popular imaging methods.