A Multiscale Discrete Feature Enhancement Network With Augmented Reversible Transformation for SAR Automatic Target Recognition

Abstract

Automatic target recognition based on synthetic aperture radar (SAR) has extensive applications in dynamic surveillance, modern airport management, and military decision-making. However, the natural mechanisms of SAR imaging introduce challenges such as target feature discretization, clutter interference, and significant scale variation, which hinder the performance of existing recognition networks in practical scenarios. As such, this article presents a novel network architecture: the multiscale discrete feature enhancement network with augmented reversible transformation. The proposed network consists of three core components: an augmented feature extraction (AFE) backbone, a discrete feature enhancement module (DFEM), and a Spider feature pyramid network (Spider FPN). The AFE backbone has the capability of effective target information preservation and clutter suppression with the aid of integration of augmented reversible transformations with intermediate supervision module and double subnetworks. The DFEM enhances both local and global discrete feature awareness through its two submodules: local discrete feature enhancement module and global semantic information awareness module. The Spider FPN overcomes target scale variation challenges, especially for small-scale targets, through a fusion-diffusion mechanism and the designed feature perception fusion module. The functionality of the proposed method is evaluated on three public datasets: SARDet-100 K, MSAR-1.0, and SAR-AIRcraft-1.0 of various polarizations and environmental conditions. Experimental results demonstrate that the proposed network outperforms current state-of-the-art methods in terms of average precision by the levels of 63.3%, 72.3%, and 67.4%, respectively.