Abstract: Aiming at the problem of insufficient data source and low detection efficiency in shaft surface defect detection,a real-time detection method of shaft surface defect based on integrated transfer learning was proposed.Through the transfer learning of pictures in similar fields,the problem of insufficient data samples was solved,and the dependence on artificial large-scale annotation data was reduced.Principal component analysis was applied to reduce the dimension of defects and extract key feature vectors,and the feature space of the defects was established.To simplify the repeated calculation of distance measurement between domains,the spatial locations were used to solve the source domain of transfer learning.The feature extractor of source domain images was trained,and the network weight of the feature extractor was transferred to You Only Look Once Version 3 (YOLO V3) target detection model to complete the knowledge transfer in the similar field,moreover the real-time detection model of shaft surface defects under high-speed production was established.Experiments showed that the method had high accuracy and robustness in real-time detection of shaft parts in production site.The integrated algorithm model had a positive detection rate of more than 97% for all kinds of defects,and the variance of the average accuracy had been reduced by nearly 3 times.

Key words: shaft surface defects, principal component analysis, integrated transfer learning, YOLO V3 target detection algorithm

摘要： 针对轴件表面缺陷分析过程中存在小样本和实时检测效率低的问题,提出一种集成迁移学习的轴件表面缺陷实时检测方法。首先通过相似领域图片的迁移学习,减少对人工大规模标注数据的经验性依赖,采用主成分分析法完成表面缺陷的降维和关键特征向量提取,建立轴件表面缺陷的特征空间,并利用空间位置求解迁移学习的源领域,降低领域间距离度量的复杂度;其次通过训练源领域图片的特征提取器,将特征提取器的网络权值迁移至YOLO V3目标检测模型中,完成相似领域的知识迁移,建立高速生产状态下的轴件表面缺陷实时检测模型。试验表明,该方法在轴件生产现场的实时检测中具有较高的准确度和鲁棒性,集成后的算法模型各类缺陷正检率达97%以上,平均精度均值的方差值缩小近3倍。

关键词: 轴件表面缺陷, 主成分分析, 集成迁移学习, YOLO V3目标检测算法