摘要
为改善人工分拣烟丝效率低下,分类效果差导致香烟品质难以得到保证的现象,提出一种将机器视觉和深度学习相结合的烟丝类型识别方法。实验采集了4种烟丝图像并经过预处理以后送入以VGG16为基础改进的Light-VGG网络模型中进行分类,改进包括减少VGG16中卷积核个数以优化网络结构;增加残差模块以提升模型学习能力;使用全局池化代替全连接层,大幅减少网络参数量,应对网络过拟合。Light-VGG相比VGG16参数量减少96.5%,预测时间减少20.3%,在自建烟丝数据集中准确率达到95.5%,也明显高于其他神经网络(AlexNet、VGG13、GoogLeNet),实现了快速、准确识别烟丝类型的目标。
In order to improve the phenomenon that the quality of cigarettes is difficult to be guaranteed due to the low efficiency of manual sorting of tobacco strands and poor classification effect, In the experiment a method for identifying tobacco strands type that combines machinevision and deep learning is proposed.In the experiment, four types of tobacco strands images were collected and preprocessing, and then sent to the improved Light-VGG network model based on VGG16 for classification. The improvements include reducing the number of convolution kernels in VGG16 to optimize the network structure;increasing the residual module to improve the learning ability of the model;using global pooling to replace the fully connected layer, greatly reduce the amount of network parameters and deal with network overfitting.Compared with VGG16,Light-VGG,reduces the number of parameters by 96.5% the prediction time by 20.3%,andthe accuracy rate in the self-built tobacco strands data set reaches 95.5%,which is also significantly higher than other neural networks(AlexNet, VGG13,GoogLeNet) and achieving the goal of quickly and accurately identifying the type of tobacco stands.
作者
牛群峰
袁强
靳毅
王莉
刘江鹏
Niu Qunfeng;Yuan Qiang;Jin Yi;Wang Li;Liu Jiangpeng(College of Electrical Engineering,Henan University of Technology,Zhengzhou 450000,China;Anyang Cigarette Factory,China Tobacco Henan Industrial Co.,Ltd.,Anyang 455000,China)
出处
《国外电子测量技术》
北大核心
2022年第9期149-154,共6页
Foreign Electronic Measurement Technology
基金
河南中烟科技项目(A202047)资助。
关键词
烟丝识别
机器视觉
深度学习
VGG16
残差模块
tobacco strands identification
machine vision
deep learning
VGG16
residual module
