Metal Defect Detection Based on Convolutional Neural Networks and Data Augmentation

This paper explores the classification of metal surface defects, focusing on using deep learning techniques to improve automated detection and classification accuracy. Traditional detection methods have limitations such as insufficient accuracy and low efficiency, while convolutional neural network methods have gradually replaced them, making significant progress. However, issues like data scarcity, complex training processes, and overfitting remain. To address these, the paper uses the ResNet50 network combined with data augmentation techniques to expand the training dataset, enhancing the model's generalization ability and accuracy. Experimental results show that the model with data augmentation outperforms the one without it, improving validation accuracy from 95% to 96%. However, the improvement is not substantial, possibly due to image distortion during augmentation. In conclusion, the ResNet50-based metal surface defect classification model, enhanced with data augmentation, has made positive progress in improving accuracy but requires further research into specific enhancement methods to reduce distortion effects. Future work could focus on optimizing data augmentation strategies and improving the model's ability to identify rare defects.