Cas-VSwin transformer: A variant swin transformer for surface-defect detection

Surface defect detection using deep learning approaches has become a promising area of research, but the difficulty of accurately locating and segmenting various forms of defects presents a challenge for this method. Swin Transformer, as a Transformer-based model, has made significant progress in computer vision. Its performance surpasses standard CNN’s performance on most tasks, but it has drawn scant attention from industrial applications. Thus far, using CNNs for surface defect detection tends to be the most common application. To explore the extensibility of the Transformer, we seek to expand the applicability of the Swin Transformer and apply it to our task. This paper proposes an improved structure called the Variant Swin Transformer. We designed a new window shift scheme that further strengthens the feature transfer between windows and makes the framework more capable of serving as a backbone for defect detection. The overall framework named the Cas-VSwin Transformer outperformed most existing models on the private dataset we built (82.3 box AP and 80.2 mask AP). We also further verified the superiority of transfer learning in training small-scale datasets. Moreover, the proposed VSwin Transformer has a lower relative error in the quantitative analysis of the defect areas, demonstrating that the Cas-VSwin Transformer is an effective model for surface defect detection, and it has great potential for other similar industrial applications. • Instance segmentation of the deep model for surface-defect detection. • Use improved Vision Transformer for industrial applications. • Annotated more than 4000 images of metal surface defects. • The proposed model outperforms most existing models on surface-defect detection. • Fine-tune the model based on transfer learning to improve accuracy.