PSD-Fuse: Spectrum-differentiated infrared and visible image fusion for environmental perception of intelligent ships

Objectives To support environmental perception for intelligent ships in complex environments, this study aims to improve target sali-ency and edge fidelity by addressing weakened target information and oversmoothed texture details in infrared and visible image fusion. Methods A Physics-aware Spectrum-Differentiated Fuse (PSD-Fuse) network is proposed. Based on the differing frequency charac-teristics of infrared target information and visible textures, the network establishes a frequency-specific processing mecha-nism spanning feature decomposition, cross-modal interaction, hierarchical reconstruction, and output optimization. Wave-let decomposition and Laplacian residual reconstruction form a multiscale decomposition and reconstruction path. At the low-frequency bottleneck, direction-adaptive scanning aggregates global context, while cross-modal gating enables feature interaction. High-frequency features are adaptively fused through a separate path and progressively incorporated to restore image details. During training, two discriminators with modality-specific subband representations are combined with an asymmetric relaxed gradient loss to balance target-information preservation and edge-texture enhancement. Results Qualitative and quantitative experiments on M³FD, TNO, and a self-built anchorage dataset show that PSD-Fuse offers ad-vantages in information preservation, global contrast, and edge-detail representation. On M³FD, edge intensity and spatial frequency exceed the respective second-best results by 25.0% and 18.7%, although structural similarity requires further improvement. In downstream object detection experiments, the mean values of both overall detection metrics exceed those obtained using the compared fusion methods. Conclusions The proposed frequency-specific processing mechanism preserves both infrared target information and visible details, shows applicability to complex water-surface scenes, and provides useful complementary information for downstream ob-ject detection.