Deep Learning-Based Architecture for RF Frame Detection for CR Applications Using Spectrograms

This paper presents a deep learning-based architecture for radio frequency (RF) frame detection using a lightweight object detector originally intended for computer vision tasks. The proposed solution was implemented using an ADALM-PLUTO software-defined radio (SDR) and a Raspberry Pi 5 - an affordable single-board computer (SBC). First, a synthetic spectrogram dataset composed of multiple Wi-Fi and Bluetooth signals is utilized to perform transfer learning using the latest1 You-Only-Look-Once (YOLO) detection model. The training process was executed over an NVIDIA RTX 3060 GPU. Then, the trained neural network was transferred to the Raspberry Pi connected to the SDR for benchmarking using over-the-air signals captured within the 2.4 GHz band. A comparative analysis with recent RF frame detection works is presented, demonstrating that our approach features smaller complexity while providing acceptable performance in terms of different object detection performance metrics and processing time2,