| 林波,黄洪琼.基于YOLOv8-FRX的无人机小目标检测改进算法[J].高技术通讯(中文),2026,36(2):179~190 |
| 基于YOLOv8-FRX的无人机小目标检测改进算法 |
| Research on an improved YOLOv8-FRX algorithm for UAV small object detection |
| |
| DOI：10. 3772 / j. issn. 1002 - 0470. 2026. 02. 007 |
| 中文关键词: 无人机； 小目标检测； 重参数化跨阶段部分网络； 小目标增强金字塔； 加权交并比 |
| 英文关键词: unmanned aerial vehicle, small object detection, reparameterized cross stage partial network, small object enhanced pyramid, wise intersection over union |
| 基金项目: |
| 作者 | 单位 | | 林波 | （上海海事大学信息工程学院上海 201306） | | 黄洪琼 | |
|
| 摘要点击次数: 170 |
| 全文下载次数: 136 |
| 中文摘要: |
| 针对无人机拍摄图像中小目标分布密集、背景复杂以及传统只看一次(you only look once， YOLO）算法对小目标的误检和漏检率较高的问题，提出了一种改进的目标检测算法YOLOv8-FRX（feature refinement and eXtended detection，FRX）。该方法通过以下3方面改进模型性能：第一，在骨干网络中引入重参数化跨阶段部分网络（reparameterized cross stage partial network，RepCSP），并在梯度流通分支上使用重参数化卷积（reparameterized convolution，RepConv），增强特征提取能力的同时减少参数量；第二，设计小目标增强金字塔（small object enhance pyramid，SOEP），提升小目标特征的捕获效率，同时保证计算效率；第三，采用加权交并比（wise intersection over union v3，Wise-IoU v3）策略，优化梯度增益分配，提高边界框回归的准确性。在VisDrone数据集上的实验结果显示，该方法的mAP50值相比基线模型提高了4.0%，参数量减少了3.6%。在无人机目标检测和跟踪数据集（unmanned aerial vehicle detection and tracking，UAVDT）上的泛化实验中，检测精度提升了1.7%，验证了该方法的通用性与有效性。上述改进不仅提升了模型对小目标的检测能力，还在一定程度上平衡了精度与效率，为无人机图像中的目标检测提供了更优解决方案。 |
| 英文摘要: |
| To address the issues of dense small object distribution, complex backgrounds, and high false detection and missed detection rates of small objects in traditional YOLO (you only look once) algorithms for drone-captured images, an improved object detection algorithm, YOLOv8-FRX(feature refinement and eXtended detection), is proposed. This method enhances model performance through three main improvements: first, introducing reparameterised cross stage partial network (RepCSP) in the backbone network and using reparameterized convolution (RepConv) on the gradient circulation branch to enhance the feature extraction capability while reducing the number of parameters; second, designing a small object enhanced pyramid (SOEP) to improve the efficiency of small object feature capture while maintaining computational efficiency; and third, adopting the Wise-IoU v3(wise intersection over union v3) strategy to optimize gradient gain allocation and enhance the accuracy of bounding box regression. Experimental results on the VisDrone dataset demonstrate that the proposed method improves the mAP50 by 4.0% compared to the baseline model while reducing parameters by 3.6%. In generalization experiments on the UAVDT(unmanned aerial vehicle detection and tracking) dataset, detection accuracy increased by 1.7%, validating the method’s generality and effectiveness. These improvements not only enhance the detection capability for small objects but also balance accuracy and efficiency, providing a superior solution for object detection in drone-captured images. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
