基于有监督对比学习的焊缝缺陷X射线检测方法  

X-ray detection method of weld defects based on supervised contrastive learning

作　　者：李国栋 吴志生[1] 彭甫镕[2] 郝康将 昝晓亮 郭威 Li Guodong;Wu Zhisheng;Peng Furong;Hao Kangjiang;Zan Xiaoliang;Guo Wei(Taiyuan University of Science and Technology,Taiyuan 030024,China;Shanxi University,Taiyuan 030006,China)

机构地区：[1]太原科技大学,太原030024 [2]山西大学,太原030006

出　　处：《焊接》2024年第7期7-14,共8页Welding & Joining

基　　金：国家重点研发计划(2018YFA0707305);山西省重点研发计划(202102050201001);山西省基础研究计划(202203021221149)。

摘　　要：【目的】基于深度学习的表面缺陷检测算法广泛应用于表面缺陷检测。然而,在焊缝缺陷检测领域,焊缝缺陷外观特征上具有同类别样本偏差大而不同类别样本偏差小的特点,这给焊缝缺陷的有效识别带来了挑战。【方法】为此,提出一种有监督对比学习的焊缝缺陷检测方法(SCL-DD),将有监督对比学习拓展到焊缝缺陷检测领域,通过正负样本进行有效的相似计算,使同一类别的缺陷样本在嵌入空间上更加接近,不同类别的缺陷彼此远离,降低类间偏差和跨类偏差对检测性能的不良影响。【结果】引入余弦分类器,通过计算特征编码与分类原型之间的余弦相似度,提高差异性缺陷样本的检测性能。在钢管焊缝表面缺陷数据集上验证所提出方法的性能。【结论】SCL-DD方法平均精度为96.9,优于其他深度学习网络。[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.

关 键 词：X射线 焊缝表面缺陷检测 智能制造 深度学习 

分 类 号：TG441.7[金属学及工艺—焊接]