Anchor-free Signal Detector Based on Center Frequency Axis Learning in Wideband Spectrogram

Deep learning-based object detectors have become increasingly critical in spectrogram-based wideband multi-signal detection, recognition, and time-frequency localization. Current methods primarily evolve from anchor-based detectors and keypoint-based anchor-free detectors. However, these detectors encounter limitations such as anchor mismatch and inefficient adaptation of detection paradigms to signal characteristics, leading to suboptimal performance. This letter proposes a novel anchor-free method that leverages a center frequency axis learning mechanism, which facilitates the capture of more robust features. Furthermore, a sample-weighting training strategy is proposed to recalibrate the contribution of samples with varying feature representation capabilities to the loss, thereby improving network training efficiency and performance. Experimental results demonstrate the superiority of the proposed method compared to existing state-of-the-art wideband signal detectors.