Registration Based Feature Core-Set Model for Defect Detection

Defect detection is an important issue in industrial applications. In this paper, we propose a few-shot defect detection model named registration based feature core-set model (RegiCore). The model learns the method of aligning image samples through a task-independent data set. It calculates the defect scores by comparing the registered features between the normal and the target images. The experimental results show that our model can effectively complete the task of defect detection and outperforms the comparison methods. Meanwhile, our model exhibits excellent ability to detect small defective targets.