基于Swin Transformer的两阶段织物疵点检测  

Two stage fabric defect detection based on Swin Transformer

作　　者：雷承霖 李敏[1,2,3] 王斌 LEI Chenglin;LI Min;WANG Bin(Wuhan Textile University,Wuhan,430200,China;Engineering Research Center of Hubei Province for Clothing Information,Wuhan,430200,China;Hubei Engineering Research Center of Intelligent Textile and Fashion,Wuhan,430200,China)

机构地区：[1]武汉纺织大学,湖北武汉430200 [2]湖北省服装信息化工程技术研究中心,湖北武汉430200 [3]纺织服装智能化湖北省工程研究中心,湖北武汉430200

出　　处：《棉纺织技术》2023年第2期41-47,共7页Cotton Textile Technology

基　　金：中国高校产学研创新基金(2020HYA02015)。

摘　　要：为了提高织物疵点检测的精准率,提出了一种基于Swin Transformer的两阶段织物疵点检测网络模型。首先,使用Swin Transformer替代传统的卷积神经网络来进行特征提取,以获得织物图像的分层特征;其次,使用神经网络架构搜索法来获取最佳特征融合网络,以得到准确的小尺寸疵点特征;最后,将融合后的特征送入多级区域建议网络,通过k-means选取最佳的候选框来进行疵点分类和位置回归。试验结果表明:对于结头、破洞等20种疵点,该疵点检测方法的mAP@0.5达到0.575;与标准的Cascade RCNN模型相比,该研究模型的mAP@0.5提升了38.1%。认为该研究提出的算法能够更好地识别各类织物疵点。In order to improve the precision for fabric defect detection,a two stage fabric defect detection network model based on Swin Transformer was proposed. Firstly,Swin Transformer replacing traditional convolutional neural network was applied to extract features so that layer features of fabric were obtained. Then,neural network framework search method was used to obtain optimized feature fusion network so that accurate features of small size defects could be obtained.Finally,features after fusing were sent to multilevel regional proposal network. Through k-means,optimized candidate box was selected for defects classification and position regression.Experimental results showed that mAP@0.5 of the method was reached 0.575 for 20 different types of defects including knot,hole and so on.Compared with standard Cascade RCNN model,the research model mAP@0.5 was increased 38.1%.It is considered that the algorithm proposed can better identify all kinds of fabric defects.

关 键 词：织物疵点 Swin Transformer 神经网络架构搜索 多级区域建议网络 Cascade RCNN 

分 类 号：TS101.9[轻工技术与工程—纺织工程]