基于深度学习的表面缺陷检测技术研究进展

Research progress of surface defect detection technology based on deep learning

随着计算机视觉技术的快速发展,基于深度学习的表面缺陷检测技术实现了爆发式的应用,并逐步成为了主流发展方向。基于深度学习的缺陷检测技术可以近似为计算机视觉任务中的分类、检测、分割等任务,其主要目的是找出物体表面缺陷的类别和所在位置,相较于传统图像处理方法,深度学习在特征提取能力和环境适应能力上优势明显。以缺陷数据标签类型为依据,对近年来基于深度学习的表面缺陷检测技术进行梳理划分,总结目前技术的优点与不足,重点阐述了监督学习下的三种缺陷检测方法。探讨了表面缺陷检测技术面临的小样本以及不平衡样本等关键问题:对于小样本问题目前有结构优化、数据增广、迁移学习等解决方法;针对不平衡样本问题,介绍了近年来热点的无监督、弱监督与半监督学习模型。随后介绍了常用的工业表面缺陷数据集并展现了近年来提出的算法在NEU数据集上的应用效果。最后对进一步的研究工作提出展望,希望能给缺陷检测研究提供有意义的参考。

With the rapid development of computer vision technology,the surface defect detection technology based on deep learning has achieved explosive applications and gradually become the mainstream development direction.Deep learning-based defect detection technology can be approximated as classification,detection,segmentation and other tasks in computer vision tasks,and the main purpose is to find out the category and location of object surface defects.Compared with traditional image processing methods,deep learning has obvious advantages in feature extraction ability and environmental adaptation ability.According to the type of defect data labels,the surface defect detection techniques based on deep learning in recent years were sorted out and divided,the advantages and shortcomings of current techniques were summarized,and three defect detection methods under supervised learning were highlighted.The key problems faced by surface defect detection technology such as small samples and unbalanced samples were discussed.For the problem of small samples,solutions including structural optimization,data widening,migration learning and other currently were analyzed;for the problem of unbalanced samples,the hot point in recent years such as unsupervised,weakly supervised and semi-supervised learning models were introduced.Then,the commonly used industrial surface defect datasets were introduced and the results of the proposed algorithms on NEU datasets were shown in recent years.Finally,prospects for further research work were proposed,hoping to provide meaningful references for defect detection research.