MRS-YOLO : A YOLO model for signal detection in multi-resolution spectrograms

REIHAN MAZOUZ, Picheral, Abigael Taylor, Jonathan Bosse, Chakib BELAFDIL, Sylvie Marcos

Keywords: Multi-resolution learning, Spectrogram object detection, YOLO, Time–frequency (TF) analysis, Attention mechanisms

TL;DR: MRS-YOLO fuses multi-resolution spectrograms to break the time–frequency tradeoff, boosting signal detection beyond single-resolution YOLO.

Abstract: Many real-world signals contain structures spanning multiple time–frequency (TF) scales, where short transients and long-duration patterns coexist. Standard spectrograms, based on the short-time Fourier transform, are constrained by the Heisenberg uncertainty principle, which here translates into the well-known trade-off between time and frequency resolutions. We propose MRS-YOLO, a multi-resolution extension of YOLO that processes spectrograms at complementary scales through parallel branches and fuses them with an attention block. On a challenging datasets of heterogeneous radio-frequency signals with spectral congestion, low SNR, and stealthy emissions, \MRS-YOLO achieves higher recall in low-SNR regimes and stronger classification accuracy than single-resolution baselines, demonstrating the value of explicit multi-scale representation learning in TF analysis. Code available at https://github.com/ICLRanonymous2026/MRS_YOLO_ICLR26.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 24326