Abstract: Underwater acoustic target identification requires the extraction of discriminative features from complex and limited vessel acoustic signals. Existing deep learning approaches for this task suffer from insufficient data samples and suboptimal performance due to inadequate feature representation capabilities, simplistic fusion strategies, and neglect of relevant knowledge extraction. To address these issues, this study proposes a network based on collaborative learning of time–frequency deep features. The framework employs a dual-branch time–frequency feature extraction network to learn deep representations directly from raw waveforms and spectrograms. A collaborative attention fusion module is introduced to integrate multi-view information through feature transformation and cross-attention mechanisms. Furthermore, a dual contrastive learning strategy is applied to regularize the feature space, thereby enhancing intra-class compactness and inter-class separability. Final recognition is achieved based on the cascaded feature representation. Experimental evaluations on the ShipsEar and DeepShip datasets demonstrate that the proposed method achieves accuracies of 96.62% and 80.97%, respectively, indicating superior classification performance.