Abstract
The expansion of marine economic activities and the increasing demand for maritime security have positioned drone-based aerial object detection as a crucial technology for applications such as marine environmental monitoring and maritime law enforcement. However, maritime aerial imagery remains highly challenging due extreme illumination variations, the small size and indistinct appearance of targets. This paper introduces a novel dual-domain contrastive learning framework integrated with low-light degradation enhancement to address these challenges. First, a low-light degradation perception module performs illumination equalization, improving image uniformity under adverse lighting conditions. Then, a dual-domain contrastive learning strategy aligns representations across both image and feature domains, enabling the detection network to learn more discriminative features. Additionally, a Local Feature Embedding and Global Feature Extraction Module (LEGM) is incorporated into the detection network to enhance the representation of small-scale maritime targets. Experiments on SeaDronessee and AFO datasets demonstrate the superiority of the proposed approach, achieving an improvement of 4.3% in mAP@0.5 and 1.9% in mAP@0.5:0.95 on SeaDronesSee, 1.9% in mAP@0.5 and 1.1% in mAP@0.5:0.95 on AFO. These results confirm that the proposed method delivers robust and accurate maritime object detection under complex environmental conditions and has strong potential for deployment in real-world maritime surveillance applications.