Deep learning algorithms for through-the-wall imaging
TL;DR
This study investigates deep learning-augmented inverse-scattering schemes that combine physics-based modeling with the learning capability of neural networks and shows that U-net achieves the highest accuracy, while Mask R-CNN offers competitive accuracy and removes boundary defects, making it highly effective for TWI applications.
Abstract
Fast through-the-wall imaging (TWI) by inverse scattering is highly desirable for many security and civilian applications. TWI by inverse scattering requires the solution of an ill-posed and nonlinear set of equations whose solution is attained iteratively (e.g., via a distorted Born iterative method). The iterative solution requires long execution times and often does not converge, limiting the applicability of inverse scattering to real-world TWI scenarios. To address these issues, this study investigates deep learning-augmented inverse-scattering schemes that combine physics-based modeling with the learning capability of neural networks. Using images after a few distorted Born iterations as input, three neural networks, namely a traditional convolutional neural network (CNN), U-net, and mask region-based CNN (Mask R-CNN) are leveraged to obtain final high-quality images of the scatterers behind the walls and their performance is compared. Among these three techniques, the traditional CNN regresses scatterer positions and restores dielectric profiles, while U-net segments scatterers and Mask R-CNN detects scatterers. Numerical results show that U-net achieves the highest accuracy, while Mask R-CNN offers competitive accuracy and removes boundary defects, making it highly effective for TWI applications.