TRTFL: A Transformer Based Robust Time-Frequency Localization Detector for Spectrogram with Overlapping Signals

Time-frequency localization based intelligent wideband spectrum sensing is essential for achieving precise dynamic spectrum access and management. Currently, object detection based detectors can achieve excellent time-frequency localization performance in a simple electromagnetic environment. However, the overlapping of signals in the time-frequency domain can corrupt signal features and degrade detector performance. In this paper, we propose a transformer based robust time-frequency localization (TRTFL) detector that fully extracts the features of the time-frequency domain correlation to improve its robustness in solving the aforementioned problem. Furthermore, to exploit the convolution operation for mining fine-grained features, we embed a convolutional layer with a small kernel in the transformer block. Finally, simulation results validate the advantages of TRTFL compared to existing detectors and demonstrate its robustness for overlapping signals in spectrogram.