Deep Learning Based High-Resolution Edge Detection for Microwave Imaging using a Variational Autoencoder

The reconstructed images in microwave imaging usually have low resolutions due to the ill-posedness of the nonlinear problem. However, Incorporating a priori information in image reconstruction algorithms can improve the results, dramatically. This information can be obtained by other imaging modalities like MRI. This article uses deep learning algorithms to obtain a priori information. Since machine learning algorithms excel at capturing complex nonlinear relationships, they can learn to approximate the underlying nonlinear mapping, between the measured scattered field and the edge of the objects which can be used as a priori information in microwave imaging algorithms. In this article, using a variational autoencoder and a fully-connected neural network, the object edges are reconstructed from measured scattered electromagnetic fields which can be used as a priori information. This work first compresses the image data using a variational autoencoder with a compression rate of 0.39%. Then, using the 6-layer fully-connected neural network, measured scattered electromagnetic fields are projected to the latent space, and finally, by using the decoder of variational autoencoder, the object edges are reconstructed using the Canny edge detection method.