| 李云文*,冯宇平*,安文志*,赵军*,聂国潇**.基于改进YOLOv8的无人机图像小目标检测算法[J].高技术通讯(中文),2024,34(7):765~775 |
| 基于改进YOLOv8的无人机图像小目标检测算法 |
| Small object detection algorithm for unmanned aerial vehicle image based on improved YOLOv8 |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2024. 07. 010 |
| 中文关键词: 小目标检测； YOLOv8n； 特征融合； 注意力机制 |
| 英文关键词: small object detection, YOLOv8n, feature fusion, attention mechanism |
| 基金项目: |
| 作者 | 单位 | | 李云文* | （*青岛科技大学自动化与电子工程学院青岛 266061)
（**烟台东方威思顿电气有限公司烟台 264000） | | 冯宇平* | | | 安文志* | | | 赵军* | | | 聂国潇** | |
|
| 摘要点击次数: 687 |
| 全文下载次数: 482 |
| 中文摘要: |
| 针对无人机（UAV）图像中目标尺寸小及特征信息少导致检测精度低的问题，提出改进YOLOv8n的小目标检测算法。首先，引入Wise IoU损失函数，通过动态非单调聚焦机制增强网络对普通质量锚框的关注，提高泛化能力。然后，添加小目标检测层（SODL）和双向特征金字塔网络（BiFPN）构建新的特征融合结构——Bi-SODL结构。SODL使网络更充分地捕捉小目标的浅层特征信息，BiFPN可以实现不同尺度特征层之间的信息交流和融合，提高小目标检测的准确性。最后，添加大型可选择模块(LSKBlock)注意力机制，通过空间选择机制和加权方式对输入特征进行处理，进一步提高小目标检测的性能和鲁棒性。实验结果表明，在VisDrone2019数据集上的检测精度指标P、mAP0.50和mAP0.50∶0.95分别提升6.4%、8.3%和5.2%，并且参数量降低25.78%。改进措施使得检测性能优于众多主流算法，证明了改进算法的有效性。 |
| 英文摘要: |
| Aiming at the problem of low detection accuracy caused by small object size and little feature information in unmanned aerial vehicle images, an improved small object detection algorithm of YOLOv8n is proposed. Firstly, the Wise-IoU loss function is introduced, which enhances the network’s focus on ordinary quality anchor frames through a dynamic non-monotonic focusing mechanism, improving the generalization ability of the algorithm. Secondly, in order to improve the accuracy of small object detection, a new feature fusion structure, the Bi-SODL structure, is constructed by adding a small object detection layer (SODL) and bidirectional feature pyramid network (BiFPN). SODL enables the network to capture the shallow feature information of the small object more adequately. BiFPN can achieve the information exchange and fusion between feature layers of different scales. Finally, LSKBlock attention mechanism is introduced, which processes the input features through spatial selection mechanism and weighting, further improving the performance and robustness of small object detection. The experimental results show that the detection accuracy metrics P, mAP_0.50 and mAP_0.50∶0.95 on the VisDrone2019 dataset are increased by 6.4%, 8.3% and 5.2% respectively, and the number of parameters is reduced by 25.78%. The improved measures make the detection performance better than many mainstream algorithms, which proves the effectiveness of the improved algorithm. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
