Time-Frequency Feature Fusion with Multi-Scale Attention for Underwater Acoustic Target Recognition
Abstract
Traditional convolutional neural networks (CNNs) for underwater acoustic target recognition (UATR) are limited by restricted local receptive fields, and reliance on frequency-domain inputs alone can discard temporal information present in the raw waveform. This study proposes TFMSUATR, a model that fuses a raw-waveform branch and a log-magnitude spectrogram branch through a learned adaptive gate, refines the fused representation with a multi-scale attention module with squeeze-and-excitation re-weighting, and models long-range dependencies with a Swin Transformer encoder. Under this protocol, TFMSUATR achieves an accuracy of 97.41% on the six-group ShipsEar task and 94.87% on DeepShip under identical grouping, clip length, and split. Controlled input ablations, a capacity-controlled comparison, additive-noise robustness, cross-dataset transfer, and Grad-CAM saliency analysis are reported. The results indicate that time-waveform fusion and multi-scale attention improve recognition relative to single-domain and single-scale baselines.