摘要
烟叶真空回潮是烟丝制取过程中非常重要的环节,使用周转箱装载烟叶送入回潮机时,周转箱需要保持较高的表面清洁度。传统的表面清洁度检测存在效率低、漏检率高等问题,针对上述问题,设计了一种基于改进卷积神经网络(Improved-Convolutional Neural Networks,I-CNN)的图像识别方法进行周转箱表面清洁度的检测。首先,在实验装置上通过多个工业相机获取洁净周转箱与含污垢周转箱的图像数据;然后,通过I-CNN网络,使用多连接路径,在对周转箱图像多尺度特征提取后进行原始路径的特征加权来实现融合特征提取;最后,基于获取的周转箱图像数据集,利用I-CNN方法提取周转箱污垢特征。测试结果表明:改进的I-CNN可以减少信息丢失,有效提升周转箱图像清洁度识别的准确率。
Tobacco vacuum retorting is a very important part of the tobacco making process,and when using crates to load tobacco to be fed into the retorting machine,the crates need to maintain a high surface cleanliness.The traditional surface cleanliness detection has the problems of low efficiency,high leakage rate,etc.For the above problems,an image recognition method based on Improved-Convolutional Neural Networks(I-CNN)is designed for the surface cleanliness detection of the crate.cleanliness detection.First,the image data of clean crates and dirt-containing crates are acquired by multiple industrial cameras on the experimental setup;then,the fusion feature extraction is realized by I-CNN network using multi-connected paths and feature weighting of the original paths after multi-scale feature extraction of the crate images;finally,the fusion feature extraction is realized by using I-CNN method based on the acquired crate image dataset.I-CNN method to extract crate dirt features.The results show that the improved I-CNN can reduce the information loss and effectively improve the accuracy of crate image cleanliness recognition.
作者
金立杰
吴盛威
王进峰
易东杰
李莹莹
刘迎超
JIN Lijie;WU Shengwei;WANG Jinfeng;YI Dongjie;LI Yingying;LIU Yingchao(Baoding Cigarette Factory of Hebei Baisha Tobacco Co.,Ltd.,Baoding 071000,China;Guizhou Aerospace Linquan Motor Co.,Ltd.,Guiyang 550008,China;Department of Mechanical Engineering,North China Electric Power University,Baoding 071003,China)
出处
《机械工程与自动化》
2025年第4期10-12,16,共4页
Mechanical Engineering & Automation
基金
国家自然科学基金资助项目(51301068)
河北省自然科学基金资助项目(E2019502060)
河北中烟工业有限责任公司科技项目(HBZY2023A019)。
关键词
表面清洁度检测
深度学习
改进卷积神经网络
surface cleanliness inspection
deep learning
improved convolutional neural networks
