Lightweight Multi-Scale Attention Network for HRRP Target Recognition
Year: 2026
Type: conference-paper
Abstract: High-resolution range profile (HRRP) is a one-dimensional radar signature that compactly encodes the projection of a target's scattering centres along the radar line of sight. HRRP recognition is a core task in radar automatic target recognition (ATR), widely used in air-defence surveillance, space situational awareness, and maritime monitoring. However, HRRP-based classification is restricted by three typical sensitivities: aspect, translation, and amplitude, as well as strong susceptibility to noise. Meanwhile, most existing deep learning methods adopt single-scale convolutions and fail to capture the inherent multi-scale scattering structures of complex targets. This paper conducts a progressive study from the Support Vector Machine (SVM) baseline and a basic Convolutional Neural Network (BasicCNN) to the proposed Multi-Scale Attention Network (MSA-Net). MSA-Net integrates a Multi-Branch Convolution (MBC) module with parallel kernels of 3, 7, and 15 for multi-scale feature extraction, and combines Squeeze-and-Excitation (SE) attention blocks to achieve adaptive channel recalibration. Experiments on a five-class simulated HRRP dataset show that although MSA-Net achieves slightly lower clean-data accuracy than the SVM baseline, it exhibits significantly stronger noise robustness. As noise intensity increases, MSA-Net can maintain accuracy above 80% of the clean-data level, while SVM and BasicCNN suffer from catastrophic performance degradation. Ablation studies verify the complementary effects of the SE mechanism and multi-branch convolution. The proposed model is lightweight with strong noise robustness, making it an effective and practical solution for real-world HRRP target recognition.
Language: English
Subfield: Aerospace Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204833939