基于改进YOLOv7-tiny的大尺寸导光板缺陷检测  

Defect Detection of Large-sized Light Guide Plates Based on Improved YOLOv7-tiny

作　　者：刘霞[1] 王洪玎 肖铭 龚烨飞 刘继承 李小伟 LIU Xia;WANG Hong-ding;XIAO Ming;GONG Ye-fei;LIU Ji-cheng;LI Xiao-wei(School of Physics and Electronic Engineering,Northeast Petroleum University;School of Electrical Engineering,Yancheng University of Technology;School of Electrical and Automation Engineering,Changshu University of Technology;Wuxi Woge Automation Co.,Ltd.)

机构地区：[1]东北石油大学物理与电子工程学院 [2]盐城工学院电气工程学院 [3]常熟理工学院电气与自动化工程学院 [4]无锡沃格自动化股份有限公司

出　　处：《化工自动化及仪表》2024年第6期1001-1009,1034,共10页Control and Instruments in Chemical Industry

基　　金：江苏省产学研合作项目(批准号:BY2021258)资助的课题;无锡市科技发展资金(批准号:G20212028)资助的课题。

摘　　要：针对导光板缺陷种类多、尺寸小、人工检测效率低的问题,提出一种基于改进YOLOv7-tiny的大尺寸导光板缺陷检测方法。首先,通过对导光板图像进行滑窗剪切以解决图像分辨率过大的问题;然后,对小样本缺陷使用多角度数据增强技术丰富数据量以解决样本不均衡的问题;最后,将轻量级卷积注意力模块(CBAM)整合到YOLOv7-tiny主干特征提取部分,使模型在通道和空间上对小目标缺陷的特征提取能力得到增强;选取WIoUv2损失函数计算定位损失,增强网络对困难示例的关注度,提高算法对低质量锚框的检测能力。实验结果表明,所提方法的均值平均精度为85.8%、召回率为81.3%,与原始YOLOv7-tiny相比,分别提高了5.4%和8.1%。Considering the multiple types of defects,small sizes and low manual detection efficiency of light guide plates,an improved YOLOv7-tiny-based method for detecting large-sized light guide plate's defect was proposed.In which,having the excessive image resolution solved by sliding window cropping on the image of the light guide plate;then,as for small sample defects,having multi-angle data augmentation techniques adopted to enrich data volume and solve imbalanced samples;finally,having the lightweight convolution attention module(CBAM) integrated into the YOLOv7-tiny backbone feature extraction section to enhance the model's feature extraction ability for small target defects in both channel and space.In addition,having the WIoUv2 loss function selected to calculate localization loss,enhance network's attention to difficult examples and improve algorithm's detection ability for low-quality anchor boxes.Experimental results show that,as compared to the traditional YOLOv7-tiny,the proposed method has an average accuracy of 85.8% and a recall rate of 81.3%,an improvement by 5.4% and 8.1%,respectively.

关 键 词：小目标缺陷检测 YOLOv7-tiny 多角度数据增强 特征提取 注意力机制 损失函数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]