Reconstruction of electromagnetic parameters of the scattering from overfilled cavities using a complex-valued convolutional neural network

We reconstruct the electromagnetic parameters of the scattering from overfilled cavities by developing a complex-valued convolutional neural network (CV-CNN) method. Since the scattering field and electromagnetic parameters are usually complex values, the input, output and weight parameters of this method are set to complex values in order to capture the characteristics of the electromagnetic field and parameters more accurately. The complex-valued matrix of the input layer is composed of the scattering field intensity received by the receivers outside the overfilled cavity, which is solved by the Petrov-Galerkin finite element interface method based on uniform non-body-fitted meshes. By means of the received scattering field intensity outside the overfilled cavities, the proposed CV-CNN can accurately reconstruct the electromagnetic parameters of arbitrarily shaped overfilled cavities with homogeneous, inhomogeneous and anisotropic media. The experimental results show that the reconstruction accuracy of CV-CNN is significantly higher than real-valued convolutional neural network (RV-CNN) with the same architecture. Moreover, the appropriate optimizer and filter size have a positive influence on the reconstruction effects. Our study can provide a new development in the electromagnetic parameter reconstruction of the cavity scattering.