An efficient detector for maritime search and rescue object based on unmanned aerial vehicle images

Abstract

Unmanned aerial vehicle (UAV) remote sensing has the advantages of responsive and high image resolution, which can better serve the object detection of maritime search and rescue (SAR). However, there are still some obstacles in maritime SAR object detection based on UAV images, due to the lack of samples for training and the complexity background of the maritime images. In this study, we build a maritime search and rescue target dataset (MSRTD) based on UAV images and further propose an efficient multi-category detector named Maritime Search and Rescue-You Only Look Once network (MSR-YOLO). To eliminate the influence of objects scale and shooting angle, we introduce the deformable convolution network (DCN) to modules in backbone. The Coordinated Attention (CA) is added to the neck of network to extract the powerful features. We replace the original detection head with decoupled detection head to better complete the task of object recognition and localization. Finally, we use Wise-Intersection over Union loss (WIoU) during the training to reduce the influence of the samples quality and help model converges rapidly. The experiments on MSRTD confirm that the proposed MSR-YOLO achieves precision, recall, and mean average precision (mAP) (0.5) of 90.00%, 68.52%, and 79.98% respectively. Compared with other methods on public dataset, ours also performs well and provides an effective detector model for maritime SAR object detection based on UAV images.