基于两层次低秩分解的无监督织物疵点检测方法  

Unsupervised fabric defect detection based on two-level low rank decomposition

作　　者：邓智超 邓开连[1] 张磊 刘肖燕 燕帅[1] DENG Zhichao;DENG Kailian;ZHANG Lei;LIU Xiaoyan;YAN Shuai(College of Information Science and Technology,Donghua University,Shanghai 201620,China)

机构地区：[1]东华大学信息科学与技术学院,上海201620

出　　处：《东华大学学报（自然科学版）》2022年第5期16-24,共9页Journal of Donghua University(Natural Science)

基　　金：国家自然科学基金青年基金资助项目(61903078,61901104);上海市松江区科技攻关项目(20SJKJGG4C);2021年“纺织之光”中国纺织工业联合会高等教育教学改革研究项目(2021BKJGLX140);上海市高等教育学会2021年度规划研究课题(Z2-10)

摘　　要：针对面料疵点检测算法泛化性差、无法适应面料种类动态变化的生产实际问题,提出了一种基于两层次低秩分解的无监督织物疵点检测方法。设计两层次低秩分解模型,并通过交替方向乘子算法对其进行求解,实现疵点与背景、噪声的解耦分离。设计基于疵点邻域图像隶属度相似性的深度聚类网络,提高疵点的定位精度。利用具有层次的训练方式,缓解网络难拟合复杂数据的问题,并提高疵点类别辨识的准确率。研究结果表明,使用基于两层次低秩分解的无监督织物疵点检测方法在格纹数据上训练模型,对格纹织物疵点的检测精度能达到81.5%,对平纹、点型和星型纹理织物疵点的检测精度也能分别达到86.1%、91.7%和95.2%。To solve the practical problem of the poor generalization of current fabric defect detection algorithms and the inability to adapt to the dynamic changes of fabric types, an unsupervised fabric defect detection algorithm based on two-level low rank decomposition was proposed. The two-level low rank decomposition model was designed and solved by the alternating direction method of multipliers to realize the decoupling and separation of defect from background and noise. The deep clustering network was designed to improve the accuracy of defect location, in terms of the similarity of the membership around defect area. To alleviate the problem that the network was difficult to fit due to the complicated data, a hierarchical training method was designed to improve the accuracy of defect category identification. The experimental results show that the model trained on the box-pattern fabric can achieve 81.5% accuracy for box-pattern fabric defect on it, and 86.1%, 91.7% and 95.2% accuracy for plain, dot-pattern and star-pattern fabric defects, respectively.

关 键 词：织物疵点检测 无监督目标检测 两层次低秩分解 层次聚类 深度聚类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]