【摘要】 在自动化工业生产环境中高效地完成产品质检是生产过程中的重要任务之一，提出一种基于回顾蒸馏学习的无监督工业品缺陷检测方法（Retro-KD）.首先，针对缺陷产生的未知性问题，采用无监督的方式训练蒸馏学习模型，同时，为了充分地利用蒸馏学习中的信息传递机制，利用中间层特征提取模块完善教师网络中的特征架构；其次，提出迭代信息融合模块，回顾地传递中间层信息，指导学生网络拟合正样本特征分布，放大缺陷样本差异性；再引入相似性度量（Structural Similarity,SSIM），增强教师与学生网络在图像空间中的相似度；最后，采用基于梯度变化的缺陷分割方法得到像素级的定位图.在MVTec-AD和Magnetic-Tile两个工业数据集上验证了该方法的有效性，其AUROC(Area under ROC)与ACC(Accuracy)指标分别提升了1.9%与1.3%.更多还原

【Abstract】 In an automated industrial production environment,efficiently completing product quality inspection is one of the important tasks in the production process. Aiming at the unknown problems caused by defects,this paper proposes an unsupervised industrial product defect detection method(Retro-KD) based on retrospective distillation learning. Firstly,This method uses an unsupervised method to train the distillation learning model. Meanwhile,in order to make full use of the information transmission mechanism in distillation learning,the middle-level feature extraction module is used to improve the feature architecture extracted by the teacher network. Secondly,an iterative information fusion module is proposed to retrospectively transfer the information of the middle layer,guide students to fit the positive sample feature distribution through the network,and amplify the difference in defective samples.Thirdly,the structural similarity(SSIM) index based on graph structure is introduced to enhance the similarity between teacher network and student network in the image space.Finally,the defect segmentation based on gradient change is adopted. This method is used to obtain a pixel-level location map. Two industrial datasets,MVTec-AD and Magnetic-Tile,are used to verify the effectiveness of the method. The AUROC(Area under ROC)and ACC(Accuracy)increases by 1.9% and 1.3% respectively.更多还原