【作者】 陈俊琰；

【导师】 郁崇文；

【作者基本信息】 东华大学 ， 数字化纺织工程， 2013， 博士

【摘要】 本文对基于计算机视觉技术的梳棉机棉网质量检测系统的开发方案与实现算法进行了系统地研究。传统的目光检测很大程度上取决于测试人员的主观判断,且效率低下,费工费时,因此本文探讨了棉网质量检测系统的构建方案和检测算法,实现了用计算机视觉检测代替工人检测,提高了纺织品生产中的自动化程度和效率。论文涉及了检测系统软硬件平台的构建、棉网数字图像的预处理、棉结杂质等疵点的图像分割识别技术、棉网均匀度的计算方法以及后续生条均匀度检测等主要内容。现将论文各章节内容简要介绍如下：第一章主要概述了计算机视觉及其检测系统应用现状以及对棉结、杂质、棉网质量检测的研究现状。主要有：论述了课题的背景和意义；介绍了计算机视觉的由来及其检测系统的应用发展现状；在介绍了棉网质量检测内容为棉结、杂质等疵点检测和棉网均匀度检测的基础上,着重介绍了国内外一些检测棉结、杂质和棉网均匀度的研究方法,及历史研究中存在的问题；文中还介绍了现有国外一些检测棉结、杂质的设备,国内目前在这方面与国外有一定差距。第二章主要探讨了梳棉机棉网质量视觉检测系统的构建方案。文中阐述了棉网质量检测系统的组成结构和工作流程,系统主要由图像采集、图像处理识别和用户界面三部分组成。在系统的主要硬件方面,首先介绍了线阵CCD与面阵CCD的区别,根据实际实验条件选择了适合本文的线阵扫描CCD传感器,同时给出了对应搭配线阵CCD的光学镜头；接着比较了几种常见的光源,在此基础上选择了白色条形LED线光源作为系统的照明装置；最后,阐述了为防止棉网出现褶皱现象而在喇叭口前放置了一对辅助压辊,同时介绍了本系统所选用的图像采集卡、计算机主机和显示器等。文中详细阐述了系统工作原理中对图像采集参数的设定。根据凸透镜成像和棉网幅宽,确定了线阵CCD的安装位置；通过前向照明和背向照明的采集效果对比,采用了前向照明作为本系统的照射方案；给出了由棉网输出速度计算图像采集速度的公式；讨论了光照强度、采集速度和光圈大小对背景亮度的关系,并给出了在一定光圈大小下,光照强度和采集速度的线性关系。同时,文中简要介绍了对棉网图像的检测原理及系统的操作步骤和特点。第三章主要介绍了数字化棉网图像的预处理方法。介绍了棉网图像采样和量化成离散数字图像的过程。对于近距离成像所产生的照度不匀问题,比较了减图像法和高帽变换法,结合后续杂质检测的原因,采用了在减图像的基础上加上图像背景最大列平均灰度值的方法,达到了去除照度不匀的效果；比较了直方图修正和同态滤波的图像增强效果,结合多个指标进行评判,采用直方图线性灰度变换对棉网图像进行增强,使之达到了便于后续检测特别是识别棉结的目的；对各种滤波去噪方法进行了比较,结合去噪效果和处理时间,选用了维纳滤波来去除图像中的噪点。第四章主要研究了棉网疵点检测的方法。介绍了棉网疵点的分类(棉结、杂质、云斑、破网),在分析各种疵点在图像上具有的特征基础上,确定了检测棉网疵点的思路,即根据灰度阈值来分割图像,面积阈值来识别疵点大小。详细阐述了图像阈值分割的原理和方法。对于棉结的检测,由于棉网的纹理结构较为复杂,采用全局阂值来对图像进行分割。介绍了五种阈值分割方法分割棉网图像的结果,结合面积阈值,通过对其中四种分割算法的比较和评判,选取了高斯拟合阈值法作为最优的分割算法。对于杂质和破网的检测,利用了其灰度比背景灰度相关来进行图像分割。文中用区域生长法对云斑进行了检测。对棉网疵点检测流程进行优化,提高了检测速度和精度。对300幅棉网图像中的棉结和杂质进行了识别,检测的统计结果与人眼目测对比,具有较高的正确检出率和较低的误检率,且算法运算速度快,说明了所选用的检测算法是合理有效的。文中还用实验对比了与AFIS棉结仪的检测结果,由于本文检测方式和原理与AFIS棉结仪的不同,得到了检测结果低度相关的结论。第五章主要研究了棉网均匀度计算方法,并由此来检测梳棉半制品生条的均匀度。在介绍了棉网均匀度定义及其数字特征的基础上,采用了区域灰度差异法和灰度共生矩阵法来定量计算棉网均匀度,与目测做对比,结果表明了区域灰度差异法中的CV(%)值能较好地评判棉网均匀度,可作为棉网均匀度的表征指标,并给出了棉网均匀度参考等级标准。根据棉网在喇叭口汇聚的实际情况,建立了棉网与生条条干之间的几何关系,并提出了计算机检测生条均匀度的方法。实验表明,计算机检测生条均匀度结果与YG135G条干仪和切段称重法的检测结果线性相关,计算机检测结果可用来表征实际生条不匀。文中还讨论了给棉罗拉和道夫的速度变化对生条不匀的影响。实验表明,检测结果符合实际生产的规律,且检测结果对梳棉工艺参数的优化起了一定的参考作用。第六章对全文进行了总结,归纳了本文的研究内容和研究结果,指出了本文的不足之处,并对今后的研究工作提出看法与展望。通过研制梳棉棉网质量计算机视觉检测系统,提供了能够识别棉网中的棉结、杂质等疵点的方法,并实现了对棉网均匀度和后续生条均匀度的检测,检测结果准确率较高。而且系统在算法设计上考虑了实时性的要求,为进一步对棉网质量的实时检测和梳棉机的实时控制打下了一定的基础。更多还原

【Abstract】 This research focused on the computer vision detecting system of the carded web quality on the cotton carding machine, including the detection and algorithm based on computer. Traditionally, the web quality is subjective visually evaluated by tester or operator, with low efficiency and long period. In this research, the web quality objective detect system is established and the corresponding algorithm is developed, which can evaluating the quality of carded web fast and conveniently, therefore, the efficiency and the degree of automatic in textile processing can be improved. In this research, the establishing of detect system, the pre-processing of detected image of carded web, the image segmentation and recognition on neps and trashes in web, and the calculation of evenness of the web and the resultant carded sliver are all dealt with. The contents of the dissertation are shown as follows:Chapter1is the general introduction and literatures review, the background of this research, the development of the visual detect system based on computer, and the previous research works focused on detecting on carded web, especially the research works in the detect on nep, trash in the web and the web evenness were reviewed and commented, some application of the detect and the devices were also introduced and commented.Chapter2focused on the establishing of computer vision detect system of carded web quality on a pilot carded machine. The construction and schematic of the system are introduced. The system is composed of three parts:image collection, image recognition processing and the operate interface of user. In the hardware of the system, the difference between linear array CCD and plane array CCD was compared and analyzed, therefore, the linear array CCD sensor was employed according to the requirement of this research, and the corresponding lens was selected. Then, some commonly used light sources were compared and analyzed, on the base of that, the LED with white line was selected as the light source of the system. Finally, a pair of assistant rollers was designed to amount in front of trumpet in card machine, which is helpful to avoid the crease of the web when it is condensed into a sliver. The parameters used in this system were introduced and optimized, the position of the CCD camera was determined on the width of the web and lens specification; the front light source was used in this system, after comparing the result between back light source and front light source; the speed of image acquisition was figured out on the base of web output speed; with a certain aperture, the relation between the intensity of illumination and the speed of image acquisition was obtained.Chapter3deals with the pre-processing of image. The collection of web image and its conversion into the discrete data image was introduced. With the reduce image method and top-hat transfer method, and consider the sequent trash detecting, the reduce image method combine average gray value of highest column on the background was used to overcome the unevenness of intensity of illumination; with the comparison of the reinforce result of correction of histogram and homomorphism filtering, the linear transfer of gray value was used to reinforce the image of web, which was convenient to sequent detect and nep recognition; with the comparison of some noise filtering and time consume, the Wiener filtering was employed to remove the noise in the web image.In Chapter4, the detecting of defaults in web was studied. The classification of web defaults, such as neps, trashes, cloud-like and broken part in web, were introduced. On the base of the characteristics of these defaults, the detecting method of the image was determined. The image segmented by gray threshold, and the size of flaws was detected by area threshold. For the detecting of nep, because the complex of texture, the global threshold was used to segment the image. By comparison the results of five methods used to segment the image, the Gauss fit threshold method was determined as the optimum method. For the detecting of trash and broken webs, the characteristic of their gray value correlation with that of background was used to segment the image. In this chapter, the cloud-like webs were detected by region growing method. Optimize the detecting process of the cotton web flaws, improving the detection speed and accuracy.The results of nep and trash recognition for300images compared to those of test results by people vision, the image recognition was much more faster with high correct detection rate and low false drop rate, which means the method used in this system is reasonable and liable. The test results were also compared with those of instrument, but because of the different mechanism, the detection of this system was different from those detected by AFIS.In chapter5, the calculation of the evenness of carded web was studied, therefore, the evenness of output sliver was figured out. On the base of the definition and characteristic of the evenness of web, the area gray value differential method and grey level co-occurrence matrix were used to calculate the evenness of the web, compare to the vision test on the web, this CV%value of area gray value differential method can be better to character the evenness of web, hence it not only can be used to evaluate the evenness of the web but also can be the standard level of evenness of the web. According to the web in trumpet, the relation between web and output sliver was established, therefore, the calculation of the evenness of output sliver by computer was proposed. The results calculated by computer and the results tested by YG135G instrument was linear related, that means the calculated results by computer can be used to present the evenness of sliver. In this chapter, the effects of speeds of both feed roller and doffer on the evenness of output sliver were also tested, the results showed that the calculated results meet the knowledge very well, it also means the method by this research was reasonable and feasible.Chapter6is the conclusion and prospect, the contents and results of this research was concluded, the deficiencies were also pointed out, the suggestion and for further research were proposed.Generally, this research established the computer vision detect system on carded web quality, the method to recognize the neps and trashes in web, the test was fast and conveniently and the results was much accuracy, which offers the base of online test and real-time control in card machine.更多还原