基于改进YOLOv3的PVC皮革瑕疵检方法研究  

Research on Defect Detection Method of PVC Leather Based on Improved YOLOv3

作　　者：唐勇 王美林[1] TANG Yong;WANG Meilin(School of Information Engineering,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学信息工程学院,广东广州510006

出　　处：《佳木斯大学学报（自然科学版）》2022年第2期119-122,共4页Journal of Jiamusi University：Natural Science Edition

基　　金：国家基金-广东省联合基金(U1701266);广东省知识产权大数据重点实验室(2018B030322016)。

摘　　要：针对目前PVC皮革瑕疵检测的检测系统中,由于皮革背景花纹的复杂性,在较多的场景中存在检测精度较低下、实时性较差等问题,提出了一种基于改进YOLOv3(You Only Look Once)的PVC皮革瑕疵检测网络YOLO-D(YOLO for Defects)。在YOLOv3特征提取网络的结构基础上,加入MixConv(Mixed Depthwise Convolutional Kernels)混合深度卷积核,以提升精度;通过引入GIoU(Generalized Intersection over Union)损失改进目标边框损失;并且加入标签平滑(Label smoothing),防止过拟合,提高模型的泛化性。在进行PVC皮革瑕疵检测中,YOLO-D在RTX-2070的显卡下平均每张检测时间为18.9ms,相比YOLOv3慢了4.8ms,但检测精度(mAP)为68.5%,相对YOLV3提高了10.9%。YOLO-D在满足PVC皮革瑕疵实时检测的情况下,较大幅度提高了检测的准确性。In the current PVC leather defect detection system,due to the complexity of leather background pattern,there are some problems such as low detection accuracy and poor real-time performance in many scenes,a PVC leather defect detection network YOLO-D(YOLO for Defects)based on improved YOLOv3(You Only Look Once)is proposed.Based on the structure of YOLOv3feature extraction network,MixConv(Mixed Depthwise Convolutional Kernels)mixed depth convolution kernel is added to improve the accuracy.Improve the target frame loss by introducing GIoU(Generalized Intersection over Union)loss.Label smoothing is added to prevent over fitting and improve the generalization of the model.In PVC leather defect detection,the average detection time of YOLO-D under RTX-2070graphics card is 18.9ms,which is 4.8ms slower than YOLOv3,but the detection accuracy(mAP)is 68.5%,which is 10.9%higher than YOLOv3.YOLO-D greatly improves the detection accuracy when it meets the real-time detection of PVC leather defects.

关 键 词：皮革瑕疵检测 YOLOv3 GIoU 混合深度卷积核 标签平滑 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]