基于Faster R-CNN的小样本表面缺陷检测算法

In recent years, the surface defect characteristics of various components in the industry have become more and more complex and diverse, and the traditional digital image processing algorithms have been unable to meet the requirements of defect detection. Most methods of object detection based on deep convolutional networks often depend heavily on the quality and quantity of training data. However in actual industrial application scenarios, qualified defect samples are often scarce. Considering this situation, this paper proposes a few-shot object detection network, based on a large number of fully annotated base class data and a small number of novel class data, using the episodic training method to improve the network\'s meta-learning ability, so as to achieve the fitting ability of small samples. In this paper, it is assumed that each class features present a K-Modal distribution in the embedded space, and a meta learning branch is added on the basis of Faster R-CNN. Experiments show that the performance of our model on DAGM 2007 dataset is significantly improved.