MES-YOLO: An efficient lightweight maritime search and rescue object detection algorithm with improved feature fusion pyramid network

Abstract

Maritime Search and Rescue (SAR) object detection is challenged by environmental complexity, variability in object scales, and real-time computation constraints of Unmanned Aerial Vehicles (UAVs). Our MES-YOLO algorithm, designed for maritime UAV imagery, employs an innovative Multi Asymptotic Feature Pyramid Network (MAFPN) to enhance detection accuracy across scales. It integrates an Efficient Module (EMO) and Inverted Residual Mobile Blocks (iRMB) to maintain a lightweight model while enhancing key information perception.The SIoU loss function is used to optimize the detection performance of the model. Tests on the SeaDronesSee dataset show that MES-YOLO increased average precision (mAP50) from 81.5% to 87.1%, reduced parameter count by 43.3%, and improved the F1 score by 6.8%, with a model size only 58.3% that of YOLOv8, surpassing YOLO series and other mainstream algorithms in robustness to background illumination and imaging angles.