| 陈志旺* **,王婷* **,吕昌昊***,张汉锋* **,彭勇****.基于多尺度融合的无人机航拍小目标检测研究[J].高技术通讯(中文),2026,36(8):849~861 |
| 基于多尺度融合的无人机航拍小目标检测研究 |
| Research on small target detection for UAV aerial images based on multi-scale fusion |
| |
| DOI：10. 3772 / j. issn. 1002 - 0470. 2026. 08. 007 |
| 中文关键词: 无人机航拍图像； 小目标检测； 多尺度融合； 可变形卷积； 最小点距离的损失函数； 可视化分析 |
| 英文关键词: unmanned aerial vehicle aerial images， small object detection， multi-scale integration， deformable convolution， minimum point distance intersection over union， visual analysis |
| 基金项目: |
| 作者 | 单位 | | 陈志旺* ** | (*燕山大学 智能控制系统与智能装备教育部工程研究中心秦皇岛 066004)
(**燕山大学 河北省工业计算机控制工程重点实验室秦皇岛 066004)
(***燕山大学 燕山大学电力电子节能与传动控制河北省重点实验室秦皇岛 066004)
(****燕山大学 电气工程学院秦皇岛 066004) | | 王婷* ** | | | 吕昌昊*** | | | 张汉锋* ** | | | 彭勇**** | |
|
| 摘要点击次数: 18 |
| 全文下载次数: 18 |
| 中文摘要: |
| 针对无人机航拍图像中小目标检测精度较低的问题，本文提出一种改进的小目标检测算法。首先，该算法在主干网络中引入可变形卷积，通过自适应调整卷积操作中的采样位置，提升对形态不规则小目标的特征提取能力；其次，在颈部网络增加一个小目标检测层，保留更多浅层细节特征，加强不同尺度特征层之间的信息融合；最后，使用基于最小点距离的损失函数(minimum point distance based IoU，MPDIoU)改进边界回归损失函数。实验部分，本文在VisDrone 2019数据集和MS COCO数据集上进行了全面验证。实验结果表明，改进后的算法在VisDrone 2019数据集上的mAP@0.50达到了39.8%，较原算法提升了2.3%；在MS COCO数据集的小目标检测平均精度均值(mean average precision，mAP)为22.5%，提升了2.8%。此外，本文还对实际场景图像和VisDrone 2019数据集的典型图像进行测试与可视化，进一步验证了所提算法的有效性。 |
| 英文摘要: |
| To address the problem of low detection accuracy of small target in unmanned aerial vehicle(UAV) aerial images, an improved small target detection algorithm is proposed in this paper. First, the algorithm introduces deformable convolution into the backbone network, and enhances the feature extraction ability of irregular small targets by adaptively adjusting the sampling position in the convolution operation. Secondly, a small target detection layer is added to the neck network to retain more shallow-layer detail features and strengthen the information fusion between different scale feature layers. Finally, the bounding box regression loss function is optimized by adopting MPDIoU(minimum point distance intersection over union). Extensive experiments are conducted on the VisDrone 2019 and MS COCO datasets. The experimental results show that the mAP@0.50 of the improved algorithm on the VisDrone2019 dataset reaches 39.8%, which is 2.3% higher than that of the original algorithm, and the small target detection mAP (mean average precision) on the MS COCO dataset is 22.5%, which is 2.8% higher than that of the original algorithm. Furthermore, visualizations of both real-world scenarios and VisDrone 2019 samples are provided, with qualitative analyses validating the algorithm’s effectiveness. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
