Adaptive Cross Transformer With Contrastive Learning for Surface Defect Detection

The presence of surface irregularities poses a significant threat to the quality of industrial products. Vision-based surface defect detection, known for its objectivity and stability, is extensively studied. Yet, accurately locating and discerning diverse defects proves challenging due to data scarcity and the diversity of defect types. To address these issues, we propose a new adaptive cross transformer with self-supervised contrastive learning, namely, ACViT-SCL, for surface defect detection. In ACViT, the cross transformer, as a model based on the Transformer architecture, is leveraged to address data scarcity issues through metalearning pipeline. Furthermore, the adaptive cross transformer is proposed to enhance the generalization of the cross Transformer across various defect detection tasks. Finally, the self-supervised contrastive learning (CL) is incorporated to enhance feature distinctiveness, fortifying resilience against diverse defects. To demonstrate the superiority and robustness of the proposed method, the performance comparison between ACViT-SCL and state-of-the-art methods is conducted on three surface defect datasets. The results demonstrate that ACViT-SCL outperforms competing methods in terms of accuracy and generalization ability.