Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (13): 209-218.DOI: 10.3778/j.issn.1002-8331.2401-0474

• Graphics and Image Processing • Previous Articles     Next Articles

Improved YOLOv8 Aerial Small Target Detection Method:CRP-YOLO

ZHAO Zhihong, HAO Ziye   

  1. 1.School of Information Science and Technology, Shijiazhuang Tiedao University, Shijiazhuang 050043, China
    2.State Key Lab of Mechanical Behavior and System Safety of Traffic Engineering Structures, Shijiazhuang Tiedao University, Shijiazhuang 050043, China
  • Online:2024-07-01 Published:2024-07-01

改进YOLOv8的航拍小目标检测方法:CRP-YOLO

赵志宏,郝子晔   

  1. 1.石家庄铁道大学 信息科学与技术学院,石家庄 050043
    2.石家庄铁道大学 省部共建交通工程结构力学行为与系统安全国家重点实验室,石家庄 050043

Abstract: UAV aerial target detection is a research hotspot in recent years. Due to the serious occlusion of small target images in the perspective of UAV aerial photography, problems such as missed detection and false detection occur. Aiming at the above problems, an improved YOLOv8 small target detection method in aerial photography is proposed: CRP-YOLO. Firstly, in order to improve the feature extraction ability of the neck network PANet, a multi-branch partial atrous convolution structure is proposed. The RFB module is combined with PConv to improve the feature fusion method of the neck network and increase the receptive field of the neck network. Secondly, the contextual Transformer (CoT) structure is introduced into C2f before the SPPF layer of the backbone network to improve the Bottleneck block, and the global context information is used to improve the feature extraction ability of the network. Finally, a small target detection head with the size of 160×160 is added to the detection layer to improve the detection ability of small targets. Experiments are carried out on the public dataset VisDrone2019. The results show that compared with the baseline model YOLOv8s, CRP-YOLO increases by 3.8 percentage points on mAP@0.5, 1.7 percentage points on mAP@0.5:0.95, and reduces the number of parameters by 1.5 MB. Compared with other mainstream target detection methods, it also obtains better detection performance.

Key words: small target detection, YOLOv8s, receptive field block (RFB), contextual Transformer (CoT)

摘要: 无人机航拍目标检测是近些年研究的热点,由于无人机航拍视角下的小目标图像及被遮挡情况严重,导致出现漏检、误检等问题。针对以上问题,提出了一种改进YOLOv8的航拍小目标检测方法:CRP-YOLO。为提升颈部网络PANet的特征提取能力,提出一种多分支部分空洞卷积结构,将RFB模块与PConv结合改进颈部网络的特征融合方式,增大颈部网络的感受野;在主干网络SPPF层前的C2f中引入CoT(contextual Transformer)结构改进Bottleneck块,利用全局上下文信息,提升网络特征提取能力;在检测层增加一个尺寸为160×160的小目标检测头,提高对小目标的检测能力。在公开数据集VisDrone2019上进行实验,结果表明,相较于基线模型YOLOv8s,CRP-YOLO在mAP@0.5上提升3.8个百分点,mAP@0.5:0.95提升1.7个百分点,参数量降低1.5?MB,与其他主流目标检测方法相比也得到较好的检测性能。

关键词: 小目标检测, YOLOv8s, 感受野模块(RFB), CoT