CNN-based hot-rolled steel strip surface defects classification: a comparative study between different pre-trained CNN models
Abdelmalek Bouguettaya,Hafed Zarzour
DOI: https://doi.org/10.1007/s00170-024-13341-0
IF: 3.563
2024-03-13
The International Journal of Advanced Manufacturing Technology
Abstract:During the manufacturing process, hot-rolled steel strip surface defects occur frequently. These defects cause economic losses and risks in the use of these products. Therefore, it is crucial to develop automatic inspection systems to identify these defects. In the last few years, computer vision has emerged as an effective tool to identify these defects. Deep learning-based computer vision techniques, especially Convolutional Neural Networks (CNNs), achieved state-of-the-art results for most computer vision tasks, including image classification. These results are obtained using a large amount of data. However, collecting such large datasets in the manufacturing field remains a challenging task. To overcome such a problem, a transfer learning-based framework with multiple CNN variants is proposed in this study. Therefore, different state-of-the-art and widely used pre-trained CNN architectures, including VGG-16, VGG-19, ResNet50, ResNet50V2, InceptionV3, InceptionResNet-V2, MobileNet-V1, MobileNet-V2, MobileNet-V3 Small, and NASNetMobile, combined with transfer learning were investigated to evaluate their performances in classifying hot-rolled steel strips surface defects. The experimental results showed that MobileNet-V2-based and InceptionResNet-V2-based methods achieve better performance than all other models in terms of accuracy, loss, training and inference times, and model size in the case of NEU-CLS dataset. Similarly, MobileNet-V1 and MobileNet-V2 provide the best performances on X-SDD dataset among the adopted models.
engineering, manufacturing,automation & control systems