Time-Frequency Image-Based Overlapping LPI Radar Signal Detection and Recognition with LPI-YOLO

Accurate recognition of overlapping low-probability-of-intercept (LPI) radar signals under low signal-to-noise ratio (SNR) conditions remains a challenging problem in electronic reconnaissance. Existing methods often struggle to separate and identify multiple overlapping components in time-frequency representations. To address this problem, we present an enhanced YOLOv8-based model, named LPI-YOLO, for processing short-time Fourier transform (STFT) spectrograms. The proposed method integrates a Coordinate Attention (CA) mechanism into the C2f backbone to improve directional feature representation along time and frequency axes. A Transformer encoder is introduced in deeper layers to model global contextual relationships among overlapping signal components. In addition, lightweight GSConv-based bottlenecks and a simplified SimSPPF module are adopted to reduce computational complexity in terms of parameters and FLOPs. Experiments on a simulated dataset containing six modulation types show that the proposed method achieves competitive performance under low SNR conditions (down to -12 dB). Compared with YOLOv8n, the proposed model shows improvements on several challenging modulation classes while maintaining lower computational complexity.