【作者】 孙娜；

【导师】 管一弘；

【作者基本信息】 昆明理工大学 ， 物理电子学， 2020， 硕士

【摘要】 随着卷烟生产技术的快速发展,卷烟的卷接和包装速度高速运行,卷烟的包装外观不可避免的出现表面异常、边缘异常、拉线异常等缺陷,严重影响了产品质量和视觉效果。目前,在生产现场通常以阈值检测方法为主,即对需要的检测区域进行比对,设定阈值进行判别。该方法操作复杂,需要对图像进行精准定位,识别效果也欠佳。为此,本文将机器视觉技术与图像处理技术应用于卷烟包装外观质量检测中,利用经典机器学习算法中的支持向量机(SVM)和BP神经网络分类模型,以及深度学习技术中的卷积神经网络模型进行研究,并进行在线测试。具体如下:本文首先对条烟图像采集和图像预处理进行研究,介绍了卷烟包装外观质量检测设备的成像系统,研究了图像预处理方法,主要包括图像去噪技术,图像锐化增强技术,条烟图像定位算法。针对条烟包装外观质量检测方法,本文首先设计了基于机器学习的检测算法,该方法通过结合小波变换和灰度共生矩阵算法提取条烟特征参数,再分别利用支持向量机和BP神经网络对条烟图像进行分类识别。实验结果表明,这两种方法的识别准确率分别为96.1%和89.9%。其次,构建了基于卷积神经网络的条烟包装外观质量检测算法,在开源框架tensorflow上设计了一种具有十个隐含层的网络结构,并通过一系列超参数实验确定合理的网络参数,确定最优性能的网络模型。实验结果表明,该网络模型在本实验数据集上的识别率达到了98.78%,识别一张图像的时间为8ms。卷积神经网络方法不需要人工提取特征,避免了传统检测方法的繁琐性,从原理上革新了传统的条烟视觉检测方法。最后,介绍了卷烟包装外观质量检测系统的结构设计,并在线安装与测试。在此之前,分析了系统的软硬件结构,并对本文研究的检测方法进行了在线测试,测试结果较好,该套条烟包装外观质量视觉检测系统满足生产检测要求。更多还原

【Abstract】 With the rapid development of cigarette production technology,cigarette welding and packaging speed running at a high speed,the packaging appearance of cigarettes inevitably has defects such as surface abnormalities,edge abnormalities,and wire drawing abnormalities,which have seriously affected product quality and visual effects.At present,the threshold detection method is usually the main method at the production site.This method compares the detection areas and sets the threshold to judge.It is complicated to operate,requires precise positioning of the image,and has poor recognition effects.In this paper,machine vision technology and image processing technology are applied to the cigarette packaging appearance quality detection.Support vector machine(SVM)and BP neural network classification model in machine learning algorithm and convolution neural network model in deep learning technology are used for research and online test.The details are as follows:This paper studies the image acquisition and image preprocessing,introduces the imaging system of the detection equipment,and studies the image preprocessing methods,mainly including image denoising technology,image sharpening enhancement technology,and cigarette image positioning algorithm.Aiming at the detection method of cigarette package appearance,this paper first designs a detection algorithm based on traditional machine learning.This method combines wavelet transform and gray level co-occurrence matrix algorithm to extract the characteristic parameters of cigarette,and then uses support vector machine and BP neural network to classify and recognize the cigarette image.The experimental results show that the recognition accuracy of the two methods is 96.1% and 89.9% respectively.Secondly,this paper proposes an algorithm based on convolution neural network to detect the appearance quality of cigarette packaging.This method designs a network structure with ten hidden layers on the open-source framework tensorflow,and determines the reasonable network parameters and the optimal performance network model through a series of hyperparameter experiments.The experimental results show that the recognition rate of the network model on the experimental data set reaches 98.78%,and the recognition time of an image is 8ms.Convolution neural network method does not need to extract features manually,which avoids the complexity of the traditional method,and innovates the traditional visual detection method.Finally,the paper introduces the structure design of cigarette packaging appearance quality detection system,and the online installation and testing.Before that,this paper analyzes the software and hardware structure of the system,and carries out online test on the detection method.The test results are good.The visual detection system of the appearance quality of the cigarette package meets the production detection requirements.更多还原