注意力残差块引导的师生网络色织物缺陷检测算法

Teacher-student Network Yarn-dyed Fabric Defect Detection Based on Attention Residual Block Guidance

摘要:

针对传统色织物缺陷检测重构模型存在难以保证缺陷区域的重构效果、漏检和误检率偏高等问题，提出一种注意力残差块引导的无监督师生网络色织物缺陷检测算法。从知识蒸馏角度出发，基于Wide_ Resnet50_2网络设计一种具有编码-解码结构的教师-学生模型，学生网络通过恢复经过预训练的教师网络的多尺度特征增强重构能力。提出一种融合双重注意力的残差模块DARM(Dual Attention Residual Module)，对特征信息进行双重权重分配的方式可以去除教师网络输出的冗余信息，进一步扩大师生网络之间对于缺陷区域的表征差异，提升模型的缺陷检测与定位能力。实验结果表明，提出的算法在YDFID-1数据集上AUPRO达到了85.8%、像素级AUROC和图像级AUROC分别达到了96.3%和98.3%；在少样本条件设置下，提出的算法在MVTec数据集上AUPRO和AUROC下降不超过4.5%，实验结果验证了该算法处理色织物缺陷检测问题的有效性以及稳定性。

Abstract:

Aiming at the traditional yarn-dyed fabric defect detection reconstruction model, there are problems such as difficult to ensure the reconstruction effect of the defective region, missed detection and high false detection rate, an unsupervised teacher-student network yarn-dyed fabric defect detection algo-rithm based on attention residual block guidance is proposed. Firstly, from the perspective of knowledge distillation, a teacher-student model with encoding-decoding structure based on Wide_Resnet50_2 network is designed, and the student network enhances the reconstruction capa-bility by recovering the multi-scale features of the pre-trained teacher network. Secondly, a DARM (Dual Attention Residual Module) is proposed to incorporate dual attention, and the dual weight assignment of feature information can remove the redundant information output from the teacher network, further expand the differences in the representation of defective regions between the teach-er-student network, and improve the defect detection and localization ability of the model. The ex-perimental results show that the proposed algorithm achieves 85.8% AUPRO, 96.3% pixel-level AUROC and 98.3% image-level AUROC on the YDFID-1 dataset, and the proposed algorithm de-creases no more than 4.5% AUPRO and AUROC on the MVTec dataset under the setting of fewer samples condition, and the experimental results validate the algorithm's effectiveness and stability in dealing with the experimental results verify the effectiveness as well as the stability of the algorithm to deal with the problem of color fabric defect detection.