A Drone Image Object Detection Algorithm Based on MSA-YOLOv10

A Drone Image Object Detection Algorithm Based on MSA-YOLOv10

Abstract—To address the problems of frequent missed detections and false detections caused by the small object scale, dense distribution, complex backgrounds, and susceptibility to occlusion in UAV aerial images, this.paper proposes MSA-YOLOv10, an improved object detection algorithm based on YOLOv10n. First, a Triple Feature Encoder (TFE) is introduced into the neck to enhance the representation of ulti-scale detail information. Second, a Scale Sequence Feature Fusion (ScalSeq) module is combined with a Channel and Position Attention Mechanism (CPAM) to improve multi-scale feature fusion and suppress background interference. Finally, a high-resolution P2 detection branch is introduced into the detection head and integrated with a Localization Quality Estimation (LQE) mechanism to enhance small-object detection and prediction-box ranking. Experiments on the VisDrone2019 dataset show that, compared with the original YOLOv10n, MSA-YOLOv10 improves Precision, Recall, and mAP50 by 1.0, 0.7, and 1.1 percentage points, respectively, with mAP50 increasing from 35.1% to 36.2%, while the number of parameters increases only from 2.27 M to 2.37 M. These results indicate that MSA-YOLOv10 has promising application value for object detection in UAV aerial imagery.

Keywords-UAV Aerial Imagery; Object Detection; Yolov10n; Multi-Scale Feature Fusion; Localization Quality Estimation; Small-Object Detection