A line spectrum detection method improved by U-net network is proposed, Which is aiming at the problem of low detection accuracy of the irregular fluctuating line spectrum in passive sonar system. The system framework is based on U-net network, and the residual structure is introduced to increase the depth of the network and strengthen the learning ability of the model feature. Meanwhile, the feature channel attention mechanism is brought in the encoder part which can help the model learn the importance of different features between channels, thus improving the feature expression ability of the model. Finally, the DUpsampling up-sampling method is used in the decoder part, and the redundancy capability in the segmented label space is utilized which can accurately restore the pixel-level prediction. The improved model is compared with the HMM model and CEM model on the line spectrum detection effect. Under the SNR of -24dB to -20dB, the accuracy of line spectrum detection of the improved model is 0.314-0.526, which is better than that of HMM model and CEM model.