计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (6): 100-109.DOI: 10.3778/j.issn.1002-8331.2311-0281

• 目标检测专题 • 上一篇    下一篇

改进YOLOv8的航拍图像小目标检测算法

付锦燚,张自嘉,孙伟,邹凯鑫   

  1. 南京信息工程大学 自动化学院,南京 210044
  • 出版日期:2024-03-15 发布日期:2024-03-15

Improved YOLOv8 Small Target Detection Algorithm in Aerial Images

FU Jinyi, ZHANG Zijia, SUN Wei, ZOU Kaixin   

  1. School of Automation, Nanjing University of Information Science & Technology, Nanjing 210044, China
  • Online:2024-03-15 Published:2024-03-15

摘要: 针对在航拍图像检测任务中,物体和整体图像尺寸都比较小,尺度特征不一和细节信息不清晰,会造成漏检和误检等问题,提出了一种改进小目标检测算法CA-YOLOv8。设计了一种通道特征部分卷积模块CFPConv(channel feature partial convolution),基于此重新构造了C2f中的Bottleneck结构,命名为CFP_C2f,从而替换YOLOv8头部和颈部的部分C2f模块,增强有效通道特征权值,提升多尺度细节特征的获取能力。嵌入一种用以提升上下文聚合能力的模块CAM(context aggregated module),优化特征通道的响应,强化对深层特征的细节感知能力。添加NWD损失函数,将其与CIoU结合作为定位回归损失函数,降低位置偏差的敏感性。充分运用多重注意力机制的优势,把原有检测头替换为DyHead(dynamic head)。在VisDrone2019数据集的实验中,改进的算法较YOLOv8s原模型参数量降低了33.3%,检测精度mAP50值和mAP50:95分别提升了8.7和5.7个百分点,表现出良好的性能,验证了其有效性。

关键词: 小目标检测, YOLOv8算法, 特征通道融合, 多重注意力

Abstract: In aerial image detection task, object and the overall image size are small, scales have different characteristics and detail information is not clear, it can cause leak and mistakenly identified problems, an improved small target detection algorithm CA-YOLOv8 is proposed. Channel feature partial convolution (CFPConv) is designed. Based on this, it reconstructs a Bottleneck structure in C2f, which is named CFP_C2f. In this way, some C2f modules in YOLOv8 head and neck are replaced, the effective channel feature weights are enhanced, and the ability to obtain multi-scale detail features is improved. A context aggregated module (CAM) is embedded to improve the context aggregation ability, optimize the response of feature channels, and strengthen the ability to perceive the details of deep features. The NWD loss function is added and combined with CIoU as a positioning regression loss function to reduce the sensitivity of position bias. By making full use of the advantages of multiple attention mechanism, the original detection head is replaced with DyHead (dynamic head). In the experiment of VisDrone2019 dataset, the improved algorithm reduces the number of parameters by 33.3% compared with the original YOLOv8s model, and the detection accuracy of mAP50 and mAP50:95 increases by 8.7 and 5.7 percentage points respectively, showing good performance and confirming its effectiveness.

Key words: small target detection, YOLOv8 algorithm, feature channel fusion, multiple attention