Small target detection method based on YOLOv8s UAV perspective

Abstract

To address the challenges in detecting objects with varying scales, dense small targets, and complex backgrounds in drone aerial imagery, the YOLOv8s model incorporates a series of targeted improvements[1]. Specifically: 1) The backbone network introduces RFAConv to overcome feature extraction bottlenecks in complex scenes; 2) The neck network is restructured using BiFPNGLSA to enhance multi-scale information fusion and spatial feature utilization; 3) A dual-layer detection architecture is introduced to specifically enhance feature representation for small objects; 4) The Inner-EIoU loss function is adopted to optimize bounding box regression accuracy. Validation on the VisDrone2019 dataset demonstrates that these enhancements comprehensively improve core metrics including precision, recall, and mAP, while simultaneously reducing model parameters. This achieves an outstanding balance between detection performance and computational efficiency.