A Mutual-Attention Guided Feature Extraction and Adaptative Decision Fusion Framework for Fine-Grained Dual-Band Radar Target Classification

Fine-grained radar target classification based on single-band, such as wideband or narrowband, poses challenges even when utilizing deep learning methods. Since different bands reflect distinct characteristics of the targets, we focus on the fine-grained classification of radar aircraft benefits from dual-band data. However, the selection of complementary dual-band features and the fusion of decisions from two bands are the key issues that need to be addressed. In order to tackle these issues, we propose a framework for fine-grained radar target classification called mutual-attention guided feature extraction and adaptative decision fusion. In this article, we propose a mutual attention selection mechanism to explore the complementary feature information between wideband and narrowband data. Furthermore, the wideband and narrowband features make distinct contributions to determine the class types. In order to address the uncertainty associated with the contribution of the wideband and narrowband data, we propose a new adaptive decision fusion strategy that adaptively assigns different weights to model the contribution uncertainty. We conducted extensive experiments on a homebrew simulated dual-band fine-grained aircraft dataset, which includes the high resolution range profile signal and the jet engine modulation signal. Compared with other classification methods, the proposed approach exhibits a remarkable classification accuracy of 95.5
$\%$
in our homebrew dataset and maintains an impressive accuracy of 87.4
$\%$
even in challenging environments with a 5 dB SNRs. Moreover, it achieves exceptional inference speeds of up to 3073 data pairs per second on the GPU: RTX3090. The results demonstrate the robustness and efficiency of the proposed method.