Transfer learning-based approach using new convolutional neural network classifier for steel surface defects classification

Abstract

Automatic surface defect detection of industrial products using visual inspection has progressively replaced manual defect detection of steel strips and become a necessary part of industrial product surface defect detection of steel strips. Various steel products exhibit a wide range of surface defects. Moreover, these defects show significant diversity and similarities, posing challenges in their classification. As a result, the models currently used for identifying these defects suffer from the challenge of low accuracy, which leaves ample opportunities for further enhancement. This paper aims to improve defect detection and classification accuracy using a new approach that combines part of a pre-trained VGG16 model as a feature extractor and a new convolutional neural network (CNN) as a classifier for classifying six types of defects appearing on steel surfaces. The experimental results have shown that our proposed method can effectively classify a variety of steel surface defects. A comparison with state-of-the-art methods shows the superiority of the proposed method.

Keywords

• Convolutional neural network
• Deep learning
• Feature extraction
• Machine learning
• Steel defects
• Transfer learning