Abstract: To enhance the radar system′s capability to resist active jamming in the complex electromagnetic environment of modern warfare, a radar compound jamming signal recognition network based on the YOLOv8 model is proposed. Initially, linear frequency modulation signals and four typical active jamming signal simulation models are established, and short-time Fourier transform (STFT) is applied for time-frequency transformation to construct a time-frequency image dataset containing various single and compound radar jamming signals. Subsequently, the impact of sample allocation ratio and the depth of the YOLOv8 model on radar jamming signal recognition effectiveness is explored, and the superiority of the YOLOv8 model is verified by comparing it with current advanced jamming recognition technologies. Finally, the robustness and generalization ability of the model are validated. Experimental results on the radar jamming signal time-frequency graph dataset demonstrate that the YOLOv8n model, with a sample allocation ratio of 0.5, achieves a recognition accuracy of 99.50% and a processing speed of 64.05 frames per second. Moreover, under low jamming-to-noise ratios within 0 dB and 20 dB the influence of unknown jamming, among the seven recognition models tested, only the YOLOv8n model still maintains an average accuracy of 89.36% and accurately recognizes the four trained jamming signals.