CPSF-Net: Confidence-Chained Position-Sensitive and Scale-Aware Fusion Network for Visible-Thermal Tiny Ship Detection

Growing demands for maritime security and navigational safety have made video surveillance indispensable for monitoring distant ship targets. Visible-thermal fusion provides a robust solution across varying weather and illumination conditions; however, tiny targets remain highly sensitive to positional errors, and cross-modal misalignment caused by occlusion, low illumination, and cluttered scenes further degrades detection accuracy. To address these issues, we propose CPSF-Net, a confidence-chained, position-sensitive and scale-aware fusion network for visible-thermal tiny ship detection. A Position-Sensitive Multimodal Fusion (PMF) module enhances and fuses pre-aligned visible-thermal features with possible residual local misalignment through cross-attention and derives a position-sensitive token-level matching confidence directly from the attention weights without introducing additional parameters. A Confidence-Guided Sparse Attention (CSA) module uses this confidence to adaptively suppress redundant cross-modal correlations and improve positional awareness. Finally, an adaptive Scale-Weighted Loss (SWL) balances Intersection over Union (IoU) and Normalized Wasserstein Distance (NWD) according to object scale and matching confidence, providing more stable localization supervision across object scales. On the AISail and RGBT-Tiny datasets, CPSF-Net achieves 92.1% and 84.2% mAP50 from unmanned surface vehicle (USV) and unmanned aerial vehicle (UAV) perspectives, respectively. These results show that the proposed method performs consistently in challenging maritime environments.