Localization of near-field sources using Vision Transformer in unknown colored noise environment

Abstract In this paper, we focus on utilizing deep learning methods for the localization of near-field signals in an unknown colored noise, which has fundamental applications in array signal processing. Specifically, a new Vision Transformer-based near-field signal localization approach (VT-NSL) is proposed by taking advantage of the self-attention mechanism of the Transformer, which circumvents the limitation of local perceptual domains in traditional convolutional neural networks. This enables VT-NSL to possess outstanding capability in capturing global information and effectively enhance the learning of critical information in colored noise environments. In contrast to typical data-driven estimation approaches that treat source localization as a multi-label task, VT-NSL takes the sample covariance matrix of the array’s received signal as input and directly outputs the localization parameters through a regression structure, which improves the training efficiency without affecting the estimation accuracy. Additionally, we explicitly derive the closed-form Cramér–Rao lower bound in the unknown colored noise environment. Extensive simulation results demonstrate the superior performance of VT-NSL compared to existing data-driven methods and traditional model-driven methods.