MSFE-YOLO: An Improved YOLOv8 Network for Object Detection on Drone View

Abstract

Due to the extremely small scales of objects in drone images, object detection on drone view is still a challenging task. To improve the accuracy of small object detection, we propose a novel object detection network multistrategy feature enhancement YOLO (MSFE-YOLO)-based YOLOv8. First, a symmetry C2f (SCF) module is constructed by expanding the symmetry feature extraction branch to enhance the feature extraction ability of backbone and neck network. Second, an efficient multiscale attention (EMA) module is used to realize cross-channel information intersection and cross-spatial learning in neck network, and it enhances the correlation of local features. Finally, a feature fusion (FF) module is designed based on SCF and EMA modules to fuse the rich low-level texture feature and the high-level semantic feature. The experimental results show that the proposed MSFE-YOLO-s achieves mAP@0.5 of 41.4% and mAP@0.5:0.95 of 25.2% (compared with YOLOv8-s, it increases by 4.1% and 3.1%, respectively) on the VisDrone2019 validation dataset and the detection rate can reach to 100 FPS on GTX 3080ti.