Multi-label Classification of Steel Surface Defects Using Transfer Learning and Vision Transformer

Steel is the most widely used metal in various industries, for instance, automotive, buildings, packaging, etc. Defect detection on steel surfaces is crucial for steel production companies. This paper proposes a multi-label classification algorithm based on deep learning methods. Several models were trained based on the Severstal dataset, namely MobileNet-V2, Xception, DenseNet121, ViT, and ResNet-50 with transfer learning. In addition, a base model is trained from scratch based on convolutional layers such as ResNet. Moreover, a ViT model is constructed according to the attention and transformer layers concept. At first, 91% for the weighted F1 score is obtained. By exploiting the weighted loss method, the weighted F1 score increased to 92%.