Sonar range profile target recognition based on spatial cross-attention mechanism for amplitude-phase fusion

Objectives To address the insufficient fusion of amplitude and phase information in active sonar range profile target recognition, an amplitude-phase dual-stream fusion recognition method preserving spatial dimensions is investigated, aiming to solve the attention degeneration problem in traditional global-concatenation strategies and the phenomenon that naive concatenation dilutes effective amplitude signals when phase quality deteriorates. Methods Logarithmic compression is applied to reduce the dynamic range of amplitude, and sine/cosine encoding is employed to eliminate phase wrapping discontinuities. Spatial-dimension-preserving dual-stream convolutional encoders are designed, upon which multi-head cross-attention is constructed to enable amplitude-phase interaction at each spatial position. Gated fusion is then introduced, which generates position-wise weights over the cross-attention-enhanced amplitude and phase features to adaptively balance their contributions. A simulated range-profile dataset based on a physical scattering model is constructed, and five-level progressive ablation experiments are conducted on both simulated and measured datasets. Results Under typical-interference simulation, cross-attention with gated fusion improves F1 score from 0.776 to 0.892 over the amplitude-only baseline, with an area under the ROC curve (AUC) of 0.974. Under strong-interference simulation, the amplitude-only baseline obtains an F1 score of 0.112, while the gated model obtains an F1 score of 0.578 and an AUC of 0.862. On the measured dataset, F1 score improves from 0.547 to 0.635 and AUC reaches 0.917, achieving the highest AUC and the smallest standard deviation (0.018) across all eight random seeds. Component ablation shows that performance decreases when positional encoding and dual pooling are enabled simultaneously and recovers after residual scaling and auxiliary classification heads are introduced. Under strong interference, cross-attention varies substantially across random seeds, whereas the gated model obtains higher F1 score and AUC. Conclusions The proposed method improves active sonar target recognition through spatial-level amplitude-phase interaction and position-adaptive gated fusion, and remains robust when phase quality deteriorates and acoustic interference increases.