Enhanced YOLOv8 Network for Small Object Detection in Drone Aerial Photography Scenarios

Image detection in drone aerial photography scenarios is crucial in both civilian and military fields. However, these images often face challenges such as low resolution, small target ratios, and complex backgrounds. To address these characteristics of drone aerial photography, we propose an improved YOLOv8 object detection network for drone aerial scenes. Firstly, to address the issue of small object feature loss caused by difficulties in multi-scale information fusion, this paper proposes the DBB-BiFPN network model. This network reduces information loss of small objects during network transmission by employing bidirectional weighted fusion of feature maps across different scales. Additionally, to address the issue of poor model generalization caused by small targets being considered low-quality samples due to their low pixel count, we have adopted Wise-IoU as the loss function for bounding box regression. This approach reduces the gradient gain between high-quality and low-quality samples, thereby enhancing the model’s localization performance and generalization ability. Furthermore, we incorporated a lightweight yet effective Triplet Attention mechanism, which facilitates information exchange across different dimensions through a three-branch structure, specifically enhancing the feature representation capability for small targets. Lastly, we introduced the Dynamic Head framework, significantly improving the target detection head’s expressive capacity without additional computational overhead. Our experiments on the VisDrone2021 dataset showed that our improved YOLOv8 algorithm achieved an AP of 41.1%, a 2.7% increase compared to the standard YOLOv8 algorithm. This robustly demonstrates the effectiveness of our method in drone aerial photography scenarios.