Research on End Face Defect Detection of Small Sample Workpiece Based on Meta Learning

In traditional recognition methods based on convolutional neural networks, the focus is on improving recognition accuracy and real-time issues, both of which depend on large- scale datasets, but ignore the issue of unbalanced datasets. This extreme imbalance in the sample set can cause the model's performance to collapse. While transfer learning can fine-tune existing pre-trained models using small-scale datasets, there is still a significant challenge in transitioning from small-scale (hundreds of sample data) to a few samples. We use the defect of car shock absorber parts as a carrier to explore intelligent few-sample recognition methods, using meta-learners to learn meta-features to achieve knowledge transfer and then guide the detector's recognition of new classes. There are still many issues to be resolved in the use of few-sample learning for practical recognition tasks. This paper proposes a generalized few-sample wheel defect recognition method based on meta-learning, using a target detector equipped with a predictor reconstruction network. The effectiveness of the method is verified.