SRT-YOLO11: a lightweight network for shortwave signal classification under low SNR conditions
Year: 2026
Type: article
Abstract: Automatic classification of shortwave communication signals at low signal-to-noise ratios (SNRs) remains challenging because discriminative time-frequency features are readily masked by strong noise, particularly in spectrum monitoring and non-cooperative communication reconnaissance. Here, we propose SRT-YOLO11, a lightweight classification model built upon the YOLO11n-cls backbone. Three complementary modules—Spatial Pyramid Pooling Fast (SPPF), RepVGG-DW, and Triplet Attention—are sequentially embedded after the C2PSA block, enhancing multi-scale time-frequency perception, local texture representation, and noise suppression via cross-dimensional recalibration, respectively. The model is evaluated on a nine-class spectrogram classification task covering typical shortwave signals and a BACKGROUND noise class for false-alarm assessment. In the reported seed-2026 run, across the SNR range of -20 to -2 dB, SRT-YOLO11 achieves an average accuracy of 36.66% and a macro-averaged F1-score of 35.13%, surpassing the baseline by 0.90 and 0.94% points, respectively, with only 0.18 M additional parameters. The ablation results quantify module interactions. Across four training seeds (2026–2029), the mean accuracy is 36.72% for SRT-YOLO11 and 35.89% for the baseline; the paired improvement is + 0.82% points (95% CI: +0.44 to + 1.20, p = 0.006). These results demonstrate that SRT-YOLO11 provides an effective and parameter-efficient solution on the evaluated synthetic AWGN benchmark; generalization to operational channels remains to be established.
Source: Scientific Reports
Institution: Xijing University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
Funder: Xijing University
OpenAlex ID: W7223108727