A YOLO-Driven Adaptive Focusing Network for Multi-Source Overlapping Modulation Perception

Abstract

Towards the challenges of high dynamic signal-to-noise ratio (SNR) variation and multi-source channel overlapping, a YOLO-driven adaptive focusing network (YOLO-AFNet) is proposed for overlapping modulation signal perception. Firstly, a time-frequency spectrum dataset of single and overlapping modulation signals is constructed. Moreover, based on adaptive multi-scale feature extraction, Focal Modulation module and region-focused feature fusion network embedded with area-attention mechanism, a lightweight YOLO-AFNet recognition model is designed. Furthermore, a Kullback-Leibler divergence-optimized adaptive random magnitude channel pruning (KL-ARMCP) algorithm is proposed for model compression, assisted by masked generative distillation to restore accuracy, enabling exceptional computational efficiency for edge deployment with limited computing resources. Additionally, the deep semantic features extracted by the pruned YOLO-AFNet are combined with random forest regression for the end-to-end SNR estimation. Experimental results demonstrate that our method achieves a high recognition accuracy of 99.28% comparable to the state-of-the-art models, with a drastic reduction of 74.39% in FLOPs and 82.39% in Params, and achieves an SNR estimation error of 0.95%. Semi-physical experimental results further validate its practicality, confirming this efficient and lightweight end-to-end visual perception paradigm as a practical solution for intelligent signal processing on resource-limited edge platforms.