We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
Development of Hybrid Models Based on AlexNet and Machine Learning Approaches for Strip Steel Surface Defect Classification.
- Authors
Boudiaf, Adel; Benlahmidi, Said; Dahane, Amine; Bouguettaya, Abdelmalek
- Abstract
The quality of the hot-rolled steel strips is essential, as they are involved in many industries, including vehicle manufacturing, electrical machines, engines, and packaging, among others. However, these products are susceptible to defects during the fabrication process and lifetime, such as cracks, holes, and scratches, leading to significant financial and reputation losses. To this end, the development of powerful AI algorithms to detect and identify possible defects in hot-rolled steel strips has gained immense attention from researchers. In the current study, we proposed various AlexNet-based models for hot-rolled steel strips' surface defect recognition, including three modified AlexNet models and 12 hybrid models. The modified AlexNet models are obtained using the transfer learning technique and replacing the classifier part with one or more new fully connected layers. On the other hand, we combined the AlexNet model, the transfer learning technique, and various machine learning algorithms like KNN, LDA, DT, and NB to obtain the hybrid models. The proposed hybrid models provide excellent classification rates, achieving an accuracy that can reach 100% in the case of Alexnet-drop7-KNN and Alexnet-drop7-DT. In addition, this work investigates the effectiveness of selecting feature extraction layers and the type of classifier on the classification accuracy rate and its training and testing time. Moreover, to further illustrate the proposed model's effectiveness, the results are compared with recent studies showing their superiority.
- Subjects
STEEL strip; SURFACE defects; ROLLED steel; CONVOLUTIONAL neural networks; K-nearest neighbor classification
- Publication
Journal of Failure Analysis & Prevention, 2024, Vol 24, Issue 3, p1376
- ISSN
1547-7029
- Publication type
Article
- DOI
10.1007/s11668-024-01927-5