【作者】 黄河；

【导师】 王学林；

【作者基本信息】 华中科技大学 ， 机械设计及理论， 2021， 硕士

【摘要】 针对一套柔性材料折叠包装设备全自动化的需要,本文研究开发了一套基于机器视觉的柔性材料折叠包装质量检测系统,能够实现对折包流程的全自动检测。本文的主要研究成果总结如下:(1)构建了折叠包装质量机器视觉检测系统。根据柔性材料折叠包装过程以及包装完成后成品的质量检测要求,在包装系统的基础上设计了PLC和PC双核、工业相机和深度相机双相机结合的包装质量检测系统方案,实现了包装过程和成品质量的自动检测与评判。(2)基于数字图像处理技术实现了对折叠包装过程质量检测。针对机器包装过程中包装材料放置位置检测,基于霍夫变换原理分别设计了基于原图与基于二值图像的定位算法,经实验验证其准确率分别达到了98.4%和100%,检测速度50ms/张;针对包装过程中折包步骤的包装完成度检测,设计了一种综合直方图匹配算法与均值哈希匹配算法的多特征融合的匹配算法,在六个步骤上进行实验平均准确率达到96.92%,检测速度87ms/张;针对包装材料破损检测、插舌回带包装材料检测,设计了一种基于像素值统计的检测算法,检测速度29ms/张时,其准确率可达100%。(3)基于深度学习算法实现了对包装成品质量检测。因包装成品正面空间形态是成品质量检测的关键指标,在原有彩色图像的基础上引入深度图像以弥补其空间信息的缺失,建立了包装成品RGB-D数据集,依托该数据集设计了不同融合时期、融合位置、后接网络、融合方式的深度图像与彩色图像融合网络,最终实验确定了一种基于Resnet18的改进的中期融合网络RGB+D＿Bottom-FC＿CAT作为检测网络,其检测准确率能达到99.141%,检测速度28ms/张。更多还原

【Abstract】 In order to realize the automatic folding packaging of flexible materials,a machine vision-based folding packaging quality detection system for flexible materials was developed in this paper,enabling the automatic detection of the folding packaging process.The main results of this paper are summarized as follow.(1)Established a machine-vision folding packaging quality detection system.According to the quality detection requirements of the folding packaging process and the final product and based on the existing flexible packaging system,we designed the quality detection system which combines the dual cores of PLC and PC as well as the dual cameras of industrial camera and depth camera,realizing the automatic detection and evaluation of the quality of the packaging process and final products.(2)Realized the quality detection of the folding packing process based on digital image processing.To detect the position of the packaging material,the original image-and binary imagebased positioning algorithms,based on the Hough transform principle,were designed.The accuracy of the two algorithms was experimentally verified and reached 98.4% and 100%,respectively,and the detection speed was 50 ms/sheet.A multi-feature fusion matching algorithm combining the histogram matching algorithm and the mean hash matching algorithm was designed to evaluate the packaging completion degree in each step during the packaging process.The average accuracy of the experiment in six steps reached 96.92%and the detection speed was 87 ms/sheet.In order to determine the damage condition and whether the tongue has brought out the packaging material,a pixel value statistics-based matching algorithm was designed.As the detection speed was 29 ms/sheet,the accuracy reached 100%.(3)Achieved the quality detection of the packed samples based on deep-learning algorithm.Given that the front morphology of the final products is the key index for their quality detection,depth image was introduced on the basis of the original color image to make up for the lack of spatial information.The RGB-D dataset of final products was established.Based on the dataset,fusion networks combining the depth image and the color image of different fusion periods,fusion positions,post networks,and fusion approaches were designed.Finally,based on Resnet18,a mid-term fusion network,RGB+D＿BottomFC＿CAT,was decided as the detection network,with the detection accuracy reaching 99.141%and the detection speed achieving 28 ms/sheet.更多还原