Fine-Tuned YOLOv26 for Automated Detection of Radio Signals in Spectrograms

This paper investigates a YOLOv26 model for automated detection of radio signals in spectrogram representations. The proposed method formulates the problem as single-class object detection in the time--frequency domain, allowing the recognition of emissions in the RF spectrum. Raw IQ data from the CSRD2025 dataset are converted into spectrograms and used to fine-tune a YOLOv26 model. Experimental results are promising, with a probability of detection of 90.87% and a precision of 97.71%. The method shows robustness in multi-signal scenarios, although performance degrades under low signal-to-noise ratio and strong overlap. These findings highlight the potential of YOLO-based models for spectrum monitoring applications.