Lightweight RadioYOLO for Radio Signal Detection

The detection of radio signals using object detection algorithms is one of the applications of deep learning in the realm of communication. However, due to the vast size of deep neural networks, inference takes longer and is more challenging to install on mobile terminals. In this paper, we present a lightweight radio detector (L-RadioYOLO), which requires fewer parameters for faster and more accurate radio signal location. Our model leverages the lightweight network GhostNet as its backbone to reduce parameters. Furthermore, YOLOV4 was used as a teacher detector to transfer knowledge to our model throughout the knowledge distillation process. We run experiments on several detectors with varying backbones, and the results show that our model achieves an Average Precision (AP) of 96.01% which is just 0.8% lower than baseline (YOLOV4), while the scale of parameters is reduced to 1/6 of baseline.