Abstract: ‍ ‍Frequency hopping signal detection in a complex electromagnetic environment is the first step for frequency hopping signal reconnaissance. This paper proposed a method called time-frequency semantic cancellation to detect signals under complex interference. This method designed a Siamese nested UNet with self-attention and graph attention mechanism. With this network， the semantic information of spectrogram containing frequency hopping signal， interference signal and noise was extracted， and then eliminated it with semantic information of the spectrogram without frequency hopping signal. Thus， the spectrogram only containing frequency hopping signal was obtained， so as to realize the signal detection. The simulation results showed that the proposed method could realize parameter estimation and blind detection of frequency hopping signals under complex interference. When signal-to-noise ratio was higher than -5 dB and signal-to-interference ratio was higher than 0 dB， the false alarm probability and missed alarm probability were less than 1‰. In time-frequency range detection， the experiment showed that the semantic cancellation detection method increased the Intersection over Union score by 0.31 compared with the semantic detection method， and the ablation experiment shows that the attention module increased the Intersection over Union score by 0.022. At the same time， the complexity of the proposed network was analyzed. The results show that this scheme has a small number of parameters and a fast-processing speed， which can be applied to the actual detection of frequency hopping signals.