Radar HRRP Sequence Target Recognition Based on a Lightweight Spatiotemporal Fusion Network

Area of Science:

• Radar Automatic Target Recognition
• Machine Learning for Signal Processing

Background:

• High-resolution range profile (HRRP) sequence recognition faces challenges like data imbalance, reduced robustness in varied conditions, and the need for lightweight models for real-time deployment on edge devices.
• Existing methods struggle to balance model efficiency with performance in complex scenarios.

Purpose of the Study:

• To propose a novel lightweight spatiotemporal fusion-based (LSTF) method for HRRP sequence target recognition.
• To address category imbalance, improve robustness, and meet real-time deployment requirements on resource-limited platforms.

Main Methods:

• Developed a lightweight Transformer encoder using group linear transformations (TGLT) for efficient temporal modeling.
• Introduced a transform-domain spatial feature extraction network combining fractional Fourier transform with an enhanced squeeze-and-excitation fully convolutional network (FSCN).
• Constructed an adaptive focal loss with label smoothing (AFL-LS) to handle class imbalance and improve generalization.

Main Results:

• The proposed LSTF method demonstrated superior performance compared to baseline approaches on MSTAR and CVDomes datasets.
• The TGLT encoder effectively reduced model size and computation for edge applications.
• The transform-domain feature extraction and AFL-LS significantly improved class separability and performance on imbalanced datasets.

Conclusions:

• The LSTF method offers an effective solution for HRRP sequence recognition, balancing lightweight design with high accuracy and robustness.
• The proposed techniques, including TGLT, fractional Fourier transform-based feature extraction, and AFL-LS, are well-suited for real-time radar automatic target recognition on edge devices.