面向遥感影像目标检测的场景关联无锚框YOLO网络OA

Object detection is one of the most crucial tasks in remote sensing image interpretation.Cur⁃rently,most deep learning-based remote sensing object detection models rely on predefined anchor boxes and often neglect contextual information in the scene,limiting detection performance and generalization ability.Based on this,this paper proposed a Scene-Related Anchor-Free YOLO(SRAF-YOLO)net⁃work tailored for remote sensing image object detection.SRAF-YOLO initially introduced a scene-en⁃hanced multi-scale feature extraction module.By fusing scene features with object features,it generated scene-enhanced features rich in contextual information.Furthermore,it utilized multi-scale operations to extract multi-scale features containing scene semantics,effectively incorporating contextual information.On this basis,a scene-assisted anchor-free detection head was designed.It utilized scene information in the feature map to constrain target class prediction,thereby enhancing detection accuracy.Simultaneously,the anchor-free structure significantly reduced the computational load associated with anchor box parame⁃ters.Experimental results on the RSOD and NWPU VHR-10 datasets demonstrate that SRAF-YOLO improves object detection accuracy by fusing scene information and utilizing the anchor-free mechanism.The mean Average Precision(mAP)reaches 94.58％and 95.95％on these datasets,respectively,mark⁃ing an improvement of 1.51％and 3.0％compared to the baseline model YOLOv8 and outperforming oth⁃er comparative methods.Validation results on external datasets further confirm the algorithm's strong gen⁃eralization ability.