基于YOLOv5的铝合金型材表面缺陷检测方法研究

Research on aluminum profile surface defects detection method based on YOLOv5

在高铁制造过程中铝合金材料占比越来越高,铝合金型材表面的质量控制和管理是生产作业中极其重要的一环。采用YOLOv5算法对铝合金型材表面缺陷检测进行实验研究。实验数据集源于阿里云天池大数据创新大赛提供的铝合金型材表面缺陷数据集,为提升样本数据的质量和数量,进行了数据增强研究。此外,网络模型训练时初始锚框的选取决定了网络反向学习的有效性,影响模型的检测精度,为此提出将K-Means++算法应用到自适应锚框算法中,解决聚类过程中初始中心点的选取问题。实验证明,经过数据增强和自适应锚框算法优化之后,YOLOv5算法表现出了良好的性能,检测速度快,同时提升了检测精度,解决了小目标以及狭长目标召回率低的问题,为高铁车身铝合金型材质量检测提供了技术支撑。

In the high-speed-railway manufacturing process,the proportion of aluminum alloy materials is getting higher and higher.The quality control and management of aluminum profile surface is extremely important.The YOLOv5 algorithm was used to detect the surface defects of aluminum profile.The experimental data set was derived from the aluminum profile surface defect data set provided by Aliyun Tianchi Competition.In order to improve the quality and quantity of sample data,the data enhancement technique was conducted.In addition,the selection of initial anchor frame during network model training determines the effectiveness of network reverse learning and affects the detection accuracy of the model.Thus,the K-Means++algorithm to applied the adaptive anchor frame algorithm was proposed to solve the problem of selecting the initial center point in the clustering process.The experiment proves that YOLOv5 algorithm has shown good performance after data enhancement and adaptive anchor frame algorithm optimization,with fast detection speed,the detection accuracy is improved,the problem of low recall rate of small target and narrow target is solved,and the technical support for aluminum profile quality detection of high-speed-railway is provided.