Abstract: ‍ ‍Accurate identification of jamming types was a prerequisite for implementing efficient anti-jamming initiatives. Aiming at the problem of low accuracy of jamming recognition under the condition of low Jamming-to-Noise Ratio（JNR）， this paper took the time-frequency images of the signal after Short-Time Fourier transform（STFT） as the training input of the convolutional neural network， proposed a neural network architecture with inverted residual structure as the main body， and introduced an attention mechanism module of joint time-frequency channel， which accurately identified the jamming type by simultaneously extracting the integrated jamming features in time-frequency and channel domains from the time-frequency images and making full use of multi-dimensional jamming feature information. The simulation results indicate that the proposed algorithm can precisely discriminate eight types of jamming signals when JNR is ‒8 dB. The recognition accuracy of eight types of jamming signals can reach more than 98.3% when JNR is ‒10 dB and can still reach more than 90% when JNR is ‒14 dB. The network complexity of the proposed algorithm was also analyzed. The outcomes indicate that the proposed scheme obtains a better compromise in time and space complexity， which verifies the superior performance of the model.