摘要
由于缺陷样本数量稀缺、缺陷种类不平衡、人工设计缺陷特征构造成本高且特征泛化能力差等实际因素,导致有监督机器学习模型难以解决色织衬衫裁片的缺陷检测问题。针对该问题,提出一种U型去噪卷积自编码器重构模型和残差分析的无监督色织衬衫裁片缺陷检测方法。首先,针对某种色织花型,使用无缺陷样本构造训练数据集;然后,建立并训练一种基于去噪U型深度卷积自编码器的重构修复模型;最后,在检测阶段,通过计算待测色织衬衫裁片图像与其重构图像之间的残差,即可快速检测和定位出缺陷区域。实验结果表明,该方法在不需要缺陷样本和缺陷标记的条件下,能有效地重构色织衬衫裁片的纹理花型,快速地检测和定位出多种色织衬衫花型的缺陷。
Due to the scarcity of defective yarn-dyed fabric samples in the textile industry,the imbalance of defect types and the high cost to manually design defect features gained the poor generalization,and the supervised model solves the problem of yarn-dyed fabric defect detection with difficulty.Therefore,an unsupervised reconstruction model is proposed based on the denoising U-shaped convolutional auto-encoder,and a residual analysis method ispresented to inspect yarn-dyed shirt piece defects.First,normal samples are collected for a specific fabric in the training phase.Second,an unsupervised reconstruction model is trained based on the denoising U-shaped deep convolutional auto-encoder,which is employed to reconstruct new test samples.Finally,calculating the residual map between the original image and correspondingly reconstructed image is used to inspect and locate areas of fabric defects.Experimental results show that the proposed method can inspect and locate many types of yarn-dyed fabric defects without any defective fabric samples.
作者
张宏伟
谭全露
陆帅
葛志强
徐健
ZHANG Hongwei;TAN Quanlu;LU Shuai;GE Zhiqiang;XU Jian(School of Electronic Information,Xi’an Polytechnic University,Xi’an 710048,China;Institute of Industrial Process Control,Zhejiang University,Hangzhou 310027,China;Institute of Engineering Medicine,Beijing Institute of Technology,Beijing 100081,China)
出处
《西安电子科技大学学报》
EI
CAS
CSCD
北大核心
2021年第3期123-130,共8页
Journal of Xidian University
基金
国家自然科学基金(61803292)
陕西省重点研发计划项目(2019ZDLGY01-08,2018GY-173)
陕西省科技厅创新人才推进计划青年科技新星项目(2018KJXX-038)
陕西省科技厅面上项目(2019JM-263)
西安工程大学研究生创新基金项目(chx2020017)。
关键词
色织衬衫裁片
缺陷检测
无监督学习
卷积自编码器
U型网络
yarn-dyed shirt piece
defect detection
unsupervised learning
convolutional auto-encoder
U net
