基于弱监督SOD网络的网状白坯织物缺陷检测方法  

Defect detection method of mesh white fabric based on weakly supervised SOD network

作　　者：刘秀平[1] 王柯欣 冯国栋 闫焕营 LIU Xiuping;WANG Kexin;FENG Guodong;YAN Huanying(Xi'an Polytechnic University,Xi'an,710048,China;Shenzhen Municipal Robotel Robot Technology Co.,Ltd.,Shenzhen,518109,China)

机构地区：[1]西安工程大学,陕西西安710048 [2]深圳罗博泰尔机器人有限公司,广东深圳518109

出　　处：《棉纺织技术》2023年第12期26-33,共8页Cotton Textile Technology

基　　金：陕西省科技厅工业领域一般项目(2018GY-173);西安市科技局项目(GXYD7.5)。

摘　　要：针对目前网状白坯织物缺陷检测过程中由于背景复杂、边界不明确等造成检测效率低的问题,提出一种基于弱监督SOD网络端到端的方法,实现对网状白坯织物疵点的检测。首先,弱监督SOD的单轮端到端训练通过草图标注实现,通过最大池化加深网络层数,减少训练时信息损失;其次,提出局部显著相干性损失和部分交叉熵损失解决草图标签不能提供详细信息的问题,并提出显著结构一致性损失,提高模型自适应性和泛化能力;最后,融合模块(CAM)聚合多层次特征得到缺陷检测结果,并引入边界细化模块,提高边界定位精度,使检测到的缺陷显著图更加清晰。利用TILDA数据集和BASLER工业相机对采集到的缺陷图像验证算法性能。试验结果表明:该研究算法精确率达到92.25%,召回率达到93.55%。认为:基于改进的弱监督SOD网络模型检测网状白坯织物的质量高且具有较好的鲁棒性。Aiming at the problems of low detection efficiency caused by complex background and unclear boundary in the current defect detection process of white mesh fabric,an end-to-end method based on weakly supervised SOD network was proposed to realize defect detection of white mesh fabric.Firstly,the single-round endto-end training of weakly supervised SOD was realized by scribble annotation,which deepened the number of network layers by maximizing pooling and reducing information loss during training.Secondly,the local significant coherence loss and partial cross entropy loss were proposed to solve the problem that sketch labels could not provide detailed information,and the significant structural consistency loss was proposed to improve the adaptability and generalization ability of the model.Finally,a fusion module(CAM)was used to synthesize multi-level features to obtain defect detection results,a boundary refinement module was introduced to improve the accuracy of boundary positioning and make the detected defect significance map clearer.The performance of the algorithm was verified by using the defect images collected by TILDA data set and BASLER industrial camera.The experiments showed that the precision of the proposed algorithm was reached 92.25%,the recall was reached 93.55%.It is considered that the improved weak-supervised SOD network model has high quality and good robustness in the detection of mesh white fabric.

关 键 词：弱监督 显著性检测 图像处理 网状白坯织物 缺陷检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]