计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (2): 202-211.DOI: 10.3778/j.issn.1002-8331.2207-0462

• 图形图像处理 • 上一篇    下一篇

改进YOLOv5的机场跑道异物目标检测算法

李小军,邓月明,陈正浩,何鑫   

  1. 1.湖南师范大学 信息科学与工程学院,长沙 410081
    2.湖南华诺星空电子技术有限公司,长沙 410221
  • 出版日期:2023-01-15 发布日期:2023-01-15

Improved YOLOv5’s Foreign Object Debris Detection Algorithm for Airport Runways

LI Xiaojun, DENG Yueming, CHEN Zhenghao, HE Xin   

  1. 1.College of Information Science and Engineering, Hunan Normal University, Changsha 410081, China
    2.Hunan NovaSky Electronic Technology Co., Ltd., Changsha 410221, China
  • Online:2023-01-15 Published:2023-01-15

摘要: 针对机场跑道异物(foreign object debris,FOD)在图像中目标占比小,特征不明显,经常导致误检、漏检的问题,提出一种改进YOLOv5的FOD目标检测算法。改进多尺度融合与检测部分,融合高分辨率特征图增强小目标特征表达,移除大目标检测层以减少网络推理计算量;引入轻量高效的卷积注意力模块(CBAM),从空间与通道两个维度提升模型关注目标特征的能力;在特征融合阶段采用RepVGG模块,提高模型特征融合能力的同时提高了检测精度;采用SIoU Loss作为损失函数,提升了边框回归的速度与精度。在自制FOD数据集上进行对比实验,结果表明:该方法在满足实时性的条件下,实现了95.01%的mAP50、55.79%的mAP50:95,比原算法YOLOv5分别提高了2.78、3.28个百分点,有效解决了传统FOD检测误检、漏检问题,同时与主流目标检测算法相比,提出的改进算法更适用于FOD检测任务。

关键词: 机场跑道异物, YOLOv5, CBAM注意力模块, RepVGG模块

Abstract: Aiming at the problem that the foreign object debris(FOD) of the airport runway has a small proportion of the target in the image and the features are not obvious, which often leads to false detection and missed detection, an improved YOLOv5 FOD target detection algorithm is proposed. Firstly, it improves the multi-scale fusion and detection part, fuses high-resolution feature maps to enhance the feature expression of small targets, and removes the large target detection layer to reduce the computational complexity of network reasoning. Secondly, a lightweight and efficient convolutional attention module(CBAM) is introduced. It improves the ability of the model to focus on target features from the two dimensions of space and channel. Then it uses the RepVGG module in the feature fusion stage to improve the feature fusion ability of the model and improve the detection accuracy. Finally, it uses SIoU Loss as the loss function to improve the speed and precision of bounding box regression. Comparative experiments are carried out on the self-made FOD data set. The results show that the method achieves 95.01% mAP50 and 55.79% mAP50∶95 under the condition of real-time performance, which is 2.78 and 3.28 percentage points higher than the original algorithm YOLOv5, respectively. It effectively solves the problems of false detection and missed detection of traditional FOD detection. At the same time, compared with mainstream target detection algorithms, the proposed improved algorithm is more suitable for FOD detection tasks.

Key words: foreign object debris(FOD), YOLOv5, CBAM attention module, RepVGG module