Unpaired Cross-Modal Contrastive Learning for Dual-Modality Ship Detection
作者
Haobin Hua,Wenxiang Xu,Kuinan Liu,Xinhai Li,J J Lu,D S Li
出处
期刊: [Institute of Electrical and Electronics Engineers] 日期:2026-05-08卷期号:: 1-6
标识
DOI:10.1109/icaisisas68969.2026.11567689
摘要
Dual-modality ship detection in port scenarios using optical and synthetic aperture radar (SAR) imagery is urgently needed. While existing research has made significant progress in optical-SAR fusion, several challenges persist: 1) Current multimodal detection models heavily rely on scarce and costly paired data with spatial alignment; 2) Modality differences, sample imbalance, and small-vessel complexity in ports further limit detection; 3) The grid-level MoE backbone lacks optimization for unpaired bimodal scenarios, hindering its use in optical-SAR ship detection. To address these issues, this paper proposes an enhanced MoE-based framework specifically tailored for unpaired optical-SAR ship detection in port scenarios. The main contributions are: 1) We introduce a 1:1 feature-level sampling mechanism to dynamically balance sample quantities between optical and SAR modalities; 2) We design a category-level contrastive loss to enhance the discriminative power of target features; 3) We implement a three-stage dynamic weight scheduling strategy to resolve optimization conflicts between contrastive learning and detection tasks, ensuring stable and efficient training. Experiments on a customized port vessel dataset Up-Det (Unpaired Detection, integrating DOTA and a subset of SARDet-100K) show that our method outperforms state-of-the-art baselines by +7.10% on SAR and +4.29% on optical modalities.