Spec-YOLO: An Efficient Deep Network for Spectrogram-Based Signals Identification

This paper introduces Spec-YOLO, a deep learning model based on You Only Look Once version 8 (YOLOv8) for automatically detecting Fifth Generation New Radio (5G NR) and Long-Term Evolution (LTE) signals in time-frequency spectrograms for next-generation wireless networks. Spec-YOLO integrates three key innovations to enhance signal identification accuracy and real-time performance: (i) selective feature fusion module enabling lightweight fusion via improved channel interaction and spatial representation; (ii) enhanced bidirectional feature pyramid network for efficiently extracting raw and positional information; and (iii) light convolution module ensuring redundant feature reduction and efficient backbone downsampling. Simulations demonstrate Spec-YOLO’s remarkable performance, by achieving a mean average precision of 92.1% at an intersection over union threshold of 0.5 (mAP0.5) and 85.5% for mAP0.95, with a compact architecture of 5.9 million parameters and 21.1 giga floating point operations per second (GFLOPs). Compared to baseline YOLOv8, Spec-YOLO represents reductions of over 47% in parameters and 26% in GFLOPs. Consequently, Spec-YOLO emerges as a highly promising, resource-efficient solution for spectrum sensing on resource-constrained mobile devices.