Abstract: Defect detection models generally require many training samples to learn the characteristics of defects, however, some important defects are often difficult to collect in the actual scene. How to learn the characteristics of these rare defects with few training samples becomes a challenging issue. To facilitate the study of few-shot defect detection, a new industrial surface defects dataset is constructed, including defect samples and defect-free samples. Meanwhile, a two-phase defect enhancement network is proposed to improve the performance of defect detection when only few samples are provided, which utilizes defect-free samples, and the whole training process is divided into two phases. The first phase needs many defect samples, while the second phase only needs few defect samples and defect-free samples. In addition, a defect prominence module is proposed, which makes better use of defect-free samples to enhance the characteristics of defective areas. Experiments on the new dataset show that the proposed defect detection model outperforms other few-shot object detection models and has a better application prospect in industrial surface defect detection.

Key words: defect detection, few-shot learning, defect-free samples, two-phase defect enhancement network, defect prominence module

摘要： 缺陷检测模型一般需要大量样本来学习缺陷的特征，但实际场景中一些重要缺陷的样本难以收集，如何用少量样本来学习罕见缺陷的特征成为一个具有挑战性的问题。为了促进少样本缺陷检测的研究，构建了一个新的工业表面缺陷数据集，包括缺陷样本和无缺陷样本。同时提出了一个两阶段缺陷增强网络以提升少样本场景下的缺陷检测性能，它利用了无缺陷样本，并将整个训练过程分为两个阶段。第一阶段的训练需要大量缺陷样本，而第二阶段的训练只需要少量缺陷样本和无缺陷样本。此外，还提出了一个缺陷突显模块，可以更好地利用无缺陷样本来增强缺陷区域的特征。在新数据集上的实验表明，该缺陷检测模型的性能优于其他的少样本目标检测模型，在工业表面缺陷检测中具有更好的应用前景。

关键词: 缺陷检测, 少样本学习, 无缺陷样本, 两阶段缺陷增强网络, 缺陷突显模块