摘要
针对烟支在生产过程中可能出现的黑点、油渍、刺破、夹沫、褶皱、缺嘴、烟支长短不一等缺陷,提出一种基于深度学习的烟支图像对比分析方法。对烟支中的水松纸区域,基于级联卷积网络,搭建了一种满足现场需求的最佳权重分布的神经网络分类模型;对烟支中的卷烟纸区域,采用高低值的模型比对算法,两者相结合大幅提高了烟支检测的实时性和准确性;同时引入了多尺度的时空特征,利用图像序列实现了帧间前后烟支缺陷信息的关联标记,将剔除轮的烟支剔除率降低了约2/3。在自建烟支数据集的基础上,搭建的分类模型准确率较ResNet-18提高了8.64个百分点,较紧固件缺陷自动检测(ADDF)算法和自动织物缺陷检测(AFDD)算法提高了7个百分点以上。
Aiming at the cigarette defects in production,such as black spots,oil stains,punctures,droplets,wrinkles,lack of mouth,and inconsistent cigarette lengths,a cigarette image comparison and analysis algorithm based on deep learning was proposed.For the contour detection and feature point regression of tipping paper in cigarettes,a neural network classification model meeting the field needs was built,making full use of the difference and correlation between samples,obtaining the optimal weight distribution;For the cigarette paper area in the cigarette,the model comparison algorithm of high and low values was adopted.The combination of the two algorithms improves the real-time and accuracy of cigarette detection.At the same time,by introducing multi-scale spatio-temporal features,the image sequence was used to realize the correlation marking of cigarette defect information before and after frames,and the cigarette rejection rate of the rejection wheel was reduced by 2/3.Based on the cigarette dataset,the accuracy of the classification model was about 8.64 percentage points higher than that of ResNet-18,and more than 7 percentage points higher than that of ADDF(Automatic Defect Detection of Fasteners)algorithm and AFDD(Automatic Fabric Defect Detection)algorithm.
作者
李学敏
谢光桥
黄卓
余楚才
LI Xuemin;XIE Guangqiao;HUANG Zhuo;YU Chucai(Mianyang Cigarette Factory,China Tobacco Sichuan Industry Limited Liability Company,Mianyang Sichuan 621000,China;Chengdu Information Technology of Chinese Academy of Sciences Company Limited,Chengdu Sichuan 610041,China)
出处
《计算机应用》
CSCD
北大核心
2023年第S01期346-350,共5页
journal of Computer Applications
关键词
级联卷积网络
时空特征
烟丝飞沫
高低值图像模板
ResNet
cascaded convolutional network
spatio-temporal feature
tobacco droplet
high and low value image template
ResNet