基于Cascade RCNN和二步聚类的织物疵点检测  

Fabric Defect Detection Based on Cascade RCNN and Two Step Clustering Method

作　　者：叶舒婷 游思晴[1] 郝灿 程智 王颖[2,3] YE Shuting;YOU Siqing;HAO Can;CHENG Zhi;WANG Ying(Beijing Wuzi University,Beijing,101149,China;Institute of Microelectronics of the Chinese Academy of Sciences,Beijing,100029,China;University of Chinese Academy of Sciences,Beijing,100049,China)

机构地区：[1]北京物资学院,北京101149 [2]中国科学院微电子研究所,北京100029 [3]中国科学院大学,北京100049

出　　处：《棉纺织技术》2022年第7期24-29,共6页Cotton Textile Technology

基　　金：北京社科基金(19GLC051);中科院STS项目(KFJSTS-QYZX-098);2021年北京市物资学院实培计划。

摘　　要：提出一种改进的基于深度卷积网络Cascade RCNN的织物疵点检测算法。针对织物存在疵点长宽比极端、疵点小以及疵点类间数量不均衡导致识别准确率低的问题,引入特征金字塔网络(FPN)和深度残差网络(ResNet101)进行高低层特征融合,获取更全面的织物疵点多尺度特征信息。采用二步聚类算法确定适用于极端形状疵点检测的预定义框最佳尺寸。采用改进的Cascade RCNN网络构架和二步聚类法确定的预定义框进行织物疵点检测试验。结果表明:改进后疵点识别准确率最高可达到98.4%。认为:改进特征提取网络和适用于极端形状疵点的预定义框能有效提高织物疵点识别准确率和定位精度。A kind of fabric defect detection algorithm based on modified deep convolutional networks Cascade RCNN was put forward.Aimed at problems of lower recognizing accuracy caused by extreme length-width ratio for defects in fabric,tiny defects and quantity imbalance among defects,characteristic feature pyramid network(FPN)and deep residual network(ResNet101)were introduced for feature fusion of high low layers.More comprehensive multiscale characteristic information of fabric defects was obtained.Two step cluster algorithm was adopted to confirm the optimized dimension of predefining frame for detection of defects in extreme shape.Modified Cascade RCNN network frame and two step clustering method were used to confirm predefining frame for fabric defect detection experiment.Test results showed that defect recognition accuracy can be reached 98.4%in maximum after modification.It is considered that modifying feature extraction network and predefining frame of defect in extreme shape can efficiently improve fabric defect recognition accuracy and positional accuracy.

关 键 词：Cascade RCNN模型 二步聚类法 织物疵点 深度残差网络 金字塔网络 预定义框 

分 类 号：TS101[轻工技术与工程—纺织工程]