251377 Enhanced Publishing丨PLS-YOLO: A Lightweight Signal Modulation Recognition Model

Serious statement: If this open source content is used in papers, books, academic reports, and other works, please cite the following referencesZhou Xiaobo, Zhang Fan *, She Chao, Zhou Guofei, Meng Jianping PLS-YOLO: A Lightweight Signal Modulation Recognition Model [J]. Journal of Electronics and Information Technology, pre published doi: 10.11999/JEIT251377Authors: Zhou Xiaobo, Zhang Fan *, She Chao, Zhou Guofei, Meng Jianping Unit: (1) School of Electronic Information Engineering, Beijing Jiaotong University (2) The First Research Institute of the Ministry of Public SecurityDOI：10.11999/JEIT250713 OriginalText: https://jeit.ac.cn/cn/article/doi/10.11999/JEIT251377 CorrespondingAuthor: Zhang Fan, [e-mail] Open source date: April 14, 2026Project Funds: National Natural Science Foundation of China General Project (52472174), National Key Research and Development Program (2023YFB3208101)Dataset Introduction: Automatic modulation recognition is a key technology for wireless communication spectrum monitoring and security assurance. This paper proposes a lightweight recognition method based on visual object detection to address the problem of high recognition accuracy and low model complexity in current deep learning based automatic modulation recognition models. Firstly, the IQ signal is transformed into a time-frequency map using short-time Fourier transform, and preprocessed using a jigsaw puzzle to transform the modulation recognition task into a visual object detection problem. Subsequently, a Precision and Lightweight Structure YOLO (PLS-YOLO) model was constructed based on YOLOv10n. This model effectively achieves a balance between recognition performance and lightweight structure by reconstructing the core modules of the network, optimizing the dimensionality reduction strategy of the backbone network channels, designing a new downsampling structure, and improving the feedforward network of the attention module. The experimental results on the RadioML2016.10a and RadioML2016.10b datasets show that the average accuracy of PLS-YOLO reaches 68.4% and 72.6%, respectively; Compared to YOLOv10n, it has reduced the number of parameters by 47.33%, decreased floating-point operations by 34.15%, and increased frame rate by 5 frames per second. The research results confirm that the PLS-YOLO model maintains excellent recognition accuracy while significantly reducing computational costs.