中国烟草学报2025，Vol.31Issue(4)：41-50,10.DOI:10.16472/j.chinatobacco.2024.T0242
基于特征增强与多尺度融合的烟包外观缺陷检测方法研究
Research on cigarette pack appearance defect detection method based on feature enhancement and multi-scale fusion
陆海华 1黄春辉 2王旭东 2曹维林1
作者信息
- 1. 浙江中烟工业有限责任公司宁波卷烟厂,浙江省宁波市奉化区葭浦西路 2001 号 315500
- 2. 厦门烟草工业有限责任公司,福建省厦门市海沧区新阳路 1 号 361004
- 折叠
摘要
Abstract
[Purpose]This study aims to solve the problems of cigarette packet appearance detection in which the target is confused with the background and the defective target is small and not easy to be recognized.[Methods]Based on improvements to YOLOv5s,a cigarette pack appearance defect detection method integrating feature enhancement and multi-scale fusion was proposed.First,a feature re-extraction module was introduced into the feature extraction network,and a combination of space-to-depth layers and non-strided convolutions was used to reduce information loss and preserve small target features.Then,a contextual attention module was introduced at the deepest layer of the feature extraction network.By learning contextual information and using deformable convolutions to extract small target features,the ability to distinguish between targets and the background was enhanced,reducing missed detections.Finally,a multi-scale receptive field enhancement module was introduced into the feature fusion network.Through a multi-branch structure,the correlation between feature information was strengthened,enhancing the semantic representation of features.[Results]Experimental results showed that FCM-YOLO achieved higher defect detection accuracy compared to other object detection algorithms.[Conclusion]The cigarette pack appearance defect detection method based on feature enhancement and multi-scale fusion effectively reduces information loss and improves the ability to distinguish between targets and the background.The detection accuracy reached 98.3%,with an FPS of 56.6,showing excellent performance,especially on easily confused categories(such as stains and damage).关键词
机器视觉/烟包外观缺陷检测/YOLOv5/多尺度融合/特征增强Key words
machine vision/inspection of appearance defects in cigarette packs/YOLOv5/multi-scale fusion/feature enhancement引用本文复制引用
陆海华,黄春辉,王旭东,曹维林..基于特征增强与多尺度融合的烟包外观缺陷检测方法研究[J].中国烟草学报,2025,31(4):41-50,10.
