结合显微视觉和注意力机制的毛羽检测方法  

Detection of yarn hairiness combining microscopic vision and attention mechanism

作　　者：唐嘉潞 杨钟亮[1] 张凇 毛新华 董庆奇 TANG Jialu;YANG Zhongliang;ZHANG Song;MAO Xinhua;DONG Qingqi(College of Mechanical Engineering,Donghua University,Shanghai 201620,China;Faculty of Science and Engineering,The University of Manchester,Manchester M139PL,UK;Beijing Chonglee Machinery Engineering Co.,Ltd.,Beijing 101111,China;Zhejiang Shuangtu New Materials Co.,Ltd.,Hangzhou 201620,China)

机构地区：[1]东华大学机械工程学院,上海201620 [2]曼彻斯特大学科学与工程学院,英国曼彻斯特M139PL [3]北京中丽制机工程技术有限公司,北京101111 [4]浙江双兔新材料有限公司,浙江杭州201620

出　　处：《智能系统学报》2022年第6期1209-1219,共11页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(51905175);浙江省健康智慧厨房系统集成重点实验室开放基金项目(2014E10014)。

摘　　要：毛羽瑕疵由于尺寸微小,极易发生漏检和误检,是化纤丝饼检测中的难点。为了准确高效地检测丝饼的外观瑕疵,搭建了基于显微视觉的毛羽瑕疵图像采集平台,提出加入注意力机制和特征融合的瑕疵检测算法(CenterNet-CBAM)。采集了毛羽和断线两类易混淆的目标图像,构建基于CenterNet-CBAM的目标检测模型,并与CenterNet、YOLOv4、Faster R-CNN、SSD四种目标检测算法进行比较。实验结果表明,工业显微相机能有效获取毛羽特征,CenterNet-CBAM在毛羽、断线两类目标的实验检测结果的准确率分别为94.00%和93.57%,召回率分别为86.75%和92.16%,AP值分别为92.93%和92.91%,两类mAP值均为92.92%,优于其他算法,验证了实验方法的有效性。Broken ends of yarns are difficult to detect due to their small features, which are easily overlooked and confused. To detect defects accurately and efficiently in the appearance of polyester yarn packages, an image acquisition platform with a microscopic camera was designed, and a defect detection algorithm(CenterNet-CBAM) with an attention mechanism and feature fusion was proposed. Two types of confusing target images, yarn hairiness, and broken ends were collected to design an objective model of CenterNet-CBAM, which was compared with four other object detection algorithms, CenterNet, YOLOv4, Faster R-CNN, and SSD. The results show that the industrial microscope camera can acquire the feature of yarn hairiness. The precision of CenterNet-CBAM for yarn hairiness and broken ends is 94.00%and 93.57%, the recall rates are 86.75% and 92.16%, AP values are 92.93% and 92.91%, respectively, and the mAP values are 92.92%. Better results in defect detection compared to other algorithms demonstrate the validity of the experiment method.

关 键 词：目标检测 显微视觉 注意力机制 化纤丝饼 瑕疵检测 照明工程 深度学习 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP41[自动化与计算机技术—计算机科学与技术]