A Multi-Task Joint Learning-Based Algorithm for Small Object Detection in Maritime UAV Imagery under Adverse Weather Conditions

In response to the challenges of adverse maritime weather in UAV-based search and rescue (SAR) missions, this paper proposes Multi-Task Sea-YOLO (MTS-YOLO), a multi-task learning detection algorithm with a dual-branch architecture. MTS-YOLO jointly optimizes object detection and denoising tasks to improve robustness. An Adaptive Feature Pyramid Fusion (AFPF) module enhances feature extraction in complex backgrounds, while an improved Multi-Scale MANet (MS-MANet) structure and a P2 prediction head strengthen small object detection. A Residual-Aware Denoising Subnet (RDS), used only during training, guides the backbone to recover semantic features from noisy inputs, boosting performance under adverse conditions. During inference, only the detection branch is retained to ensure real-time speed. Extensive experiments are conducted on the proposed WeatherSeaDronesSee dataset, which includes adverse maritime scenarios, show that MTS-YOLO achieves state-of-the-art results. It surpasses mainstream models, with a mAP50 of 73.1%, improving by 11.8% over YOLOv8n. These results highlight the model's effectiveness and deployment potential for real-world maritime UAV SAR scenarios.