Fast Array Ground Penetrating Radar Localization by CNN-Based Optimization Method

Abstract

This article presents an optimization-based approach to overcome redundancy arising from the multivariables enumeration process in multiple signal classification (MUSIC). By incorporating Broyden–Fletcher–Goldfarb–Shanno (BFGS) optimization, the computational speed of the MUSIC algorithm is significantly improved while maintaining mathematical accuracy. The optimization techniques require reasonable initial values to start the iteration, while for single target imaging purposes, the initial values can be acquired by the boundary between the near field and the far field. To generate suitable initial values for the optimization, we employ a modified convolutional neural network (CNN) to approximate the boundaries between the near and far fields, which vary with array system properties. Besides, the proposed method introduces a method for the Hessian matrix and gradient initialization for the BFGS method. Using simulation results as training samples, the modified CNN successfully establishes boundary approximations. Simulation and experimentation confirm the feasibility of our proposed method, showing its advantages in both accuracy and computation speed compared to existing techniques.

Keywords

• Broyden–Fletcher–Goldfarb–Shanno (BGFS)
• convolutional neural network (CNN)
• multiple signal classification (MUSIC)
• quasi-Newton method