TPS-YOLOv8n: an advanced algorithm for small object detection in complex maritime search and rescue scenarios using unmanned aerial vehicle imagery

Abstract Unmanned aerial vehicles (UAVs) play a crucial role in maritime search and rescue (SAR) missions due to their rapid deployment and high-resolution imaging capabilities. However, detecting small objects in dynamic maritime environments remains challenging, owing to ocean waves, low contrast and motion blur, which degrade detection performance. To address these challenges, we propose TPS-YOLOv8n, an enhanced detection model based on YOLOv8. The acronym ‘TPS’ denotes three key components: a Tiny prediction head, a poly context attention (PCA) module and soft non-maximum suppression (Soft-NMS). These modules are designed to improve small object detection in complex maritime SAR scenarios. Specifically, the tiny prediction head enhances localization on low-resolution feature maps. The PCA module integrates multi-scale convolutional kernels within the cross stage partial network fusion (C2f) structure to enhance feature representation in cluttered backgrounds. Finally, Soft-NMS is employed to suppress redundant detections and reduce false positives. Experimental results on the SeaDronesSee and AFO datasets demonstrate the effectiveness of TPS-YOLOv8n. It achieves an mAP50 of 78.1% on SeaDronesSee, surpassing YOLOv8n by 13.6% and attains 93.7% on AFO, with a 1.6% improvement. These results highlight the robustness of TPS-YOLOv8n and its potential for reliable UAV-based object detection in maritime SAR applications.