Objective To address the inadequate fusion of amplitude and phase information in active sonar range profile target recognition, this study proposes a spatially preserving, dual-stream fusion method for amplitude and phase features. The method aims to overcome two limitations of conventional global-concatenation strategies: the collapse of attention scores to scalar values, which undermines the ability of the attention mechanism to capture spatially varying relationships, and the dilution of informative amplitude features by direct concatenation when phase information is degraded.
Method Logarithmic compression is applied to reduce the dynamic range of amplitude data, while sine-cosine encoding is employed to eliminate discontinuities caused by phase wrapping. Two convolutional encoders that preserve spatial dimensions are designed to extract amplitude and phase features separately. A multi-head cross-attention mechanism is then introduced to facilitate interactions between the two feature streams at each spatial position. Subsequently, a gated fusion mechanism generates position-specific weights for the cross-attention-enhanced amplitude and phase features, enabling their contributions to be adaptively balanced. A simulated sonar range-profile dataset based on a physical scattering model is constructed, and five-stage progressive ablation experiments are conducted using both simulated and measured datasets.
Results Under typical interference conditions in simulation, the combination of cross-attention and gated fusion increases the F1 score from 0.776 for the amplitude-only baseline to 0.892, with an area under the ROC curve (AUC) of 0.974. Under strong interference, the amplitude-only baseline achieves an F1 score of 0.112, whereas the gated fusion model achieves an F1 score of 0.578 and an AUC of 0.862. On the measured dataset, the F1 score improves from 0.547 to 0.635 and the AUC reaches 0.917, achieving the highest AUC and the smallest standard deviation (0.018) across eight random seeds. Ablation studies reveal that simultaneously incorporating positional encoding and dual pooling reduces performance, which recovers after residual scaling and auxiliary classification heads are introduced. Under strong interference, the cross-attention model exhibits high variability across random seeds, whereas the gated fusion model achieves higher F1 score and AUC values.
Conclusion The proposed method improves active sonar target recognition through spatial-level interactions between amplitude and phase features and position-adaptive gated fusion. These mechanisms enhance recognition robustness when phase information degrades and acoustic interference increases.
DownLoad: