Abstract: To fully exploit the discriminative information in ship-radiated noise and enhance the robustness of passive underwater target recognition, a novel underwater target recognition framework based on time-frequency feature decoupling is proposed. Firstly, a two-stage line-spectrum modulation decomposition algorithm based on a multi-scale strategy is employed to achieve efficient separation of line-spectrum and modulation information from the received signal in the time-frequency domain. Subsequently, a time-frequency decoupling and fusion module with a dual-branch network structure is utilized to adaptively fuse the time-frequency features of the separated signals, thereby enhancing the discriminative power of the time-frequency features. The proposed method models the time-frequency distribution characteristics of ship-radiated noise during both signal preprocessing and neural network design, reducing the impact of complex marine environmental interference on the underwater target recognition model. Systematic experimental validation was conducted using various time-frequency feature extraction methods on the public DeepShip dataset and measured data from the Yellow Sea. The results demonstrate that the proposed method achieves high recognition accuracies of 80.02% and 97.81% on the two datasets, respectively, significantly outperforming existing approaches and verifying its effectiveness in practical application scenarios.