应用局部纹理特征的无监督织物瑕疵检测算法  

Unsupervised fabric defect segmentation using local texture feature

作　　者：周建[1,2] 王静安[1] 高卫东[1,2] ZHOU Jian WANG Jing＇an GAO Weidong(College of Textile and Clothing, Jiangnan University, Wuxi, Jiangsu 214122, China Key Laboratory of Eco-Textiles （Jiangnan University）, Ministry of Education, Wuxi, Jiangsu 214122, China)

机构地区：[1]江南大学纺织服装学院,江苏无锡214122 [2]生态纺织教育部重点实验室(江南大学),江苏无锡214122

出　　处：《纺织学报》2016年第12期43-48,共6页Journal of Textile Research

基　　金：国家自然科学基金项目(61501209);江苏高校优势学科建设工程资助项目(苏政办发2011-137号);江苏省博士后科研资助计划项目(1601017A)

摘　　要：针对当前算法对种类繁多瑕疵,尤其是经纬向瑕疵适应差问题,提出一种应用局部纹理特征的无监督织物瑕疵检测算法。这种算法采用无监督检测方案,检测过程中不需要参考样本。在检测过程中,首先根据瑕疵稀少性特点,直接从整体织物图像中获取表征局部织物纹理的局部二值模式直方图特征;然后利用机织物经纬交织特点对局部织物图像沿经纬向投影,并在此基础上提取特征;最后计算所提取特征的瑕疵异常图,并对其进行权重方式融合后实施阈值分割,实现瑕疵检测。实验结果表明,所提出的投影特征能有效表征局部织物纹理,与局部二值模式特征结合使用能有效检测织物瑕疵。Aiming at the poor versatility of existing methods in various fabric defect types especially for warp and weft direction defects,this work presents unsupervised fabric defect segmentation using local texture feature. The proposed algorithm adopts unsupervised scheme,without need of any reference samples. For detection,the rarity of fabric defects is used to obtain local binary pattern（ LBP） histogram features that can represent the local fabric texture from the entire image. Then,benefiting from the characteristics of woven fabrics＇ interlacing structure, and the one-dimension vectors obtained by projecting fabric image into warp and weft directions are extracted to represent local texture. Lastly,the anomaly maps of defect are computed from the extracted features,which are fused to segment defect with weight factors used. The experimental results show that the proposed projection feature along warp and weft directions can well represent local fabric texture,which can achieve satisfied results in identifying defects by combining with LBP features.

关 键 词：织物瑕疵 纹理表征 经纬向投影 异常检测 

分 类 号：TS101.9[轻工技术与工程—纺织工程]