【摘要】 随着机器视觉算法的发展与完善，各类深度学习方法逐步取代人眼检测与基于特征选择的传统计算机视觉方法，应用于工业生产的各个环节，对各类表面缺陷进行检测。深度学习方法随着网络层次深入，能由浅至深提取图像特征，但由于其基于数据驱动，需要巨量数据作为支撑，这与工业生产中缺陷异常样本数据量小，且分布不均相互矛盾。针对以上问题，基于仅包含325幅图像样本的小样本连接器数据集，在目标检测网络YOLOv5的基础上，提出一种基于权重迁移与模型调整的方法，采用冻结与解冻训练相结合的方式训练目标网络。实验表明，对于该小样本数据集，相较于直接运用目标检测网络，该方法具有更高的检测精度与更快的收敛速度，更能满足工业生产需求。更多还原

【Abstract】 With the development and improvement of machine vision algorithms in recent years, various deep learning methods have begun to gradually replace human-eye detection and traditional computer vision methods based on feature selection, and are applied in many aspects of industrial production to detect all kinds of surface defects that may occur in production. The deep learning method can extract image features from shallow to deep with the deepening of the network layer. However, because it is data-driven,it needs a huge amount of data as support, which is contradictory to the small amount of abnormal sample data and uneven distribution in industrial production. For the above problems, in this paper, using the small-sample connector dataset containing only 325 image samples, based on the object detection network YOLOv5, the method based on weight transfer and model adjustment is proposed, and the network is trained by a combination of freezing and thawing training. Experiments show that for this small-sample data set,compared with the direct use of the object detection network, this method has higher detection accuracy and faster convergence speed,which can satisfy the needs of industrial production better.更多还原