基于YOLO v7的海量烟支外观缺陷快速自动标注方法  

作　　者：吕献周 蒋铭 李庆松[1] 吴仕超 余茜 

机构地区：[1]红云红河烟草(集团)有限责任公司会泽卷烟厂,云南曲靖654299 [2]云南财经大学,昆明650000

出　　处：《科技创新与应用》2024年第15期40-45,共6页Technology Innovation and Application

基　　金：红云红河烟草(集团)有限责任公司科技项目(HYHH2022ZK01)。

摘　　要：图像标注作为监督式机器学习的关键环节,在处理海量的烟支缺陷数据时,传统的人工标注方法由于耗时长和主观性强等缺点显得不够高效。针对烟支缺陷检测领域中大量图像数据的自动化标注挑战,该文以YOLO v7作为基线网络,并进行一系列经验性的改进,以解决传统人工标注过程中存在的高成本和低效率问题。通过对YOLO v7的结构进行创新性调整,如合并neck层和head层,并引入Rep VGG结构,实现烟支图像的高效自动标注。实验结果表明,改进后的YOLO v7和YOLO v7-tiny在真实烟支数据集上的标注错误率分别为7.3%和6.56%,其中YOLO v7-tiny展现最快的标注速度。这项研究不仅在提高标注效率和准确性方面取得显著进步,还为烟支缺陷检测领域提供一种经济高效的自动化处理方案。Image tagging is a key part of supervised machine learning.When dealing with massive cigarette defect data,the traditional manual labeling method is not efficient because of its long time and strong subjectivity.Aiming at the challenge of automatic labeling of a large number of image data in the field of cigarette defect detection,this paper uses YOLO v7 as the baseline network and makes a series of empirical improvements to solve the problems of high cost and low efficiency in the traditional manual labeling process.Through the innovative adjustment of the structure of YOLO v7,such as merging neck layer and head layer,and introducing Rep VGG structure,the efficient automatic label of cigarette image is realized.The experimental results show that the labeling error rates of the improved YOLO v7 and YOLO v7-tiny on the real cigarette data set are 7.3%and 6.56%respectively,and YOLO v7-tiny shows the fastest labeling speed.This study has not only made remarkable progress in improving the efficiency and accuracy of labeling,but also provides an economical and efficient automatic processing scheme for cigarette defect detection.

关 键 词：YOLO v7 烟支外观缺陷检测 自动标注 RepConv VGG 

分 类 号：F768.29[经济管理—产业经济]