Abstract: Objective Deep learning-based surface defect detection algorithms are widely used in surface defect detection. However, in the field of weld defect detection, appearance characteristics of weld defects are characterized by a large deviation of samples of the same category and a small deviation of samples of different categories, which poses a challenge to the effective recognition of weld defects. Methods To this end, a supervised contrastive learning method for weld defect detection (SCL-DD) was proposed to extend supervised contrastive learning to the field of weld defect detection, where effective similarity computation was carried out through positive and negative samples, so that defect samples of the same class were closer to each other in the embedding space, and defects from different classes were far away from each other, and negative effects of the interclass bias and cross-class bias on detection performance were reduced. Results A cosine classifier was introduced to improve detection performance of differential defect samples by calculating the cosine similarity between feature encoding and classification prototype. Performance of the proposed method was validated on a steel pipe weld surface defect dataset. Conclusion SCL-DD method had an average accuracy of 96.9, which was better than other advanced deep learning networks.