A Lightweight YOLOv5-Based Network for Signal Modulation Recognition

Signal modulation recognition plays a significant role in communication countermeasures. However, traditional signal modulation recognition methods encounter challenges such as high computational complexity and sensitivity to signal-to-noise ratio (SNR), leading to slow recognition speeds and inaccurate results in practical applications. To address these issues, this paper proposes a lightweight signal modulation recognition network (LSMRN) with the signal time-frequency spectrogram as the representation. Specifically, the proposed network utilizes a lightweight backbone feature extraction network, MobileNetV3, to effectively reduce model complexity and improve recognition speed. The coordinate attention (CA) mechanism encodes spatial information from the time-frequency spectrogram into the channel attention module, enhancing the model's ability to recognize images with low SNR. Experimental results demonstrate that the LSMRN achieves an average recognition accuracy of 99.4% with FPS of 27.292, representing a 5.65 improvement over the YOLOv8n model.