An underwater acoustic target recognition model based on heterogeneous spectral attention feature fusion

Because of the complexity of the marine environment, underwater acoustic target recognition (UATR) faces challenges such as difficulties in feature extraction and low robustness when dealing with low signal-to-noise ratio (SNR) data. To address these issues, an underwater target recognition method based on heterogeneous spectral attention feature fusion is proposed. A multi-feature representation strategy is adopted, in which low-frequency narrowband line spectral features of underwater acoustic signals are comprehensively captured through feature extraction methods including short-time Fourier transform (STFT), Mel-spectrogram (Mel), and constant-Q transform (CQT), thereby constructing a more complete and enriched signal representation space. A multi-stage, hierarchical deep feature modeling framework is systematically designed, comprising a feature enhancement module, a feature fusion module, and a recognition module. The Time-Frequency Transformer (TF-Transformer) and convolutional neural network (CNN) block (TCB) module, which combines TF-Transformers and CNNs, are used to jointly model global and local dependencies in underwater signals, while a cross-attention mechanism is employed to achieve adaptive fusion of heterogeneous spectral features. Experimental results show that the proposed method achieves recognition accuracies of 98.26% and 95.58% on the ShipsEar and DeepShip datasets under low SNR conditions, confirming its superior recognition performance and robustness across diverse datasets and noise environments.