A FAST PULSE SIGNAL DETECTION METHOD BASED ON THE MG-YOLOv11n NETWORK

This study presents MG-YOLOv11n, a lightweight adaptation of YOLOv11n for rapid detection of pulse signals in time–frequency representations. To improve discrimination between signals with closely related spectral patterns, the network incorporates a multiscale attention module that emphasizes informative regions and feature channels. Computational demands are reduced by substituting selected conventional convolutions with ghost convolutions and removing low-contribution channels through L1-regularization-guided structured pruning. The resulting network is fine-tuned to recover detection performance after compression. Experimental evaluation indicates that MG-YOLOv11n achieves high detection accuracy while requiring fewer model parameters and less computation, supporting efficient analysis of pulse signals.