【作者】 王朝莉；

【导师】 钟跃崎； 王善元；

【作者基本信息】 东华大学 ， 数字化纺织工程， 2012， 博士

【摘要】 异形纤维面料已经成为当前功能性面料消费市场一个重要分支。它通过使用非圆形喷丝板来改变纤维的横截面形状这种最简单的改性方法来改善普通圆形截面的合成纤维的光泽、手感、蓬松性、抗污性、抗起毛起球性等特性。异形纤维横截面形状的表征和自动识别对于分析截面的非圆形化对纤维特性的影响机理、纺织品质量控制和指导纺织品性能设计具有重要意义。但是目前,合成纤维横截面形态的定量分析是通过人眼观察、近似测量来得到,现行标准中的异形度指标体系也存在一定的缺陷,例如,可能会出现异形度存在明显差异但是指标的值却相等的情况。异形纤维截面形状的识别更是要依赖于人的主观判断来实现,非常耗时、耗力。所以,急需要一种能够完成异形纤维横截面形态分析的自动检测系统。有鉴于此,本课题借助图像处理技术,在课题组已建立的自动获取纤维显微图像系统这一条件基础上,研究了适用于异形纤维显微图像的纤维横截面形态自动分析系统。建立了适用于异形纤维显微图像的图像处理系统,分析了异形纤维横截面的形状特征,对纤维异形度的指标表征和纤维形状种类的自动识别做了深入研究。提出了全新的、能够应对任意实心异形纤维截面形状的表征指标,实现了自动测量；通过边界到质心点距离的波动曲线的比较提出了异形截面的识别特征参数提取方法。并在此基础上,提出了基于特征向量之间的非相似度量的顺序聚类算法,完成了异形纤维横截面形状种类的自动识别。在异形纤维显微图像处理方面,对可能出现的各类噪声的来源和分布特点进行了分析,从纤维异形度表征和识别的应用角度分析了异形纤维显微图像处理的特殊性,设计出最适合异形纤维显微图像的处理算法和流程,大大提高了异形纤维图像二值分割的精度。其中关键的两个步骤分别为：(一)针对异形纤维显微图像的特点,提出了能够最大化增强目标(纤维的边界)和背景(纤维横截面纹理和外部包埋体)对比度的图像增强方法；(二)通过计算包含标注对象的最小凸多边形的方法,大大改善了普通阈值分割算法的精度。实验结果证明,本文提出的图像增强算法降噪效果明显,而且有效地消除了对后续阈值分割非常不利的光照不匀现象。本文提出的计算包含标注对象的最小凸多边形的阈值分割后处理方法,能够很好地保护纤维的边界完整,有助于准确而彻底地去除噪声,大大改善了普通阈值分割算法的精度。该方法与常用的边缘检测算子相比较,提取出的纤维边界更加完整、准确,消除了双边缘和各种虚假边界,同时噪声点也较少。从后续处理的角度分析,也降低了随后的特征提取的技术难度。在纤维异形度的指标表征方面,突破了以往基于内切圆和外接圆的传统表征方法,借助纤维截面的距离-角度函数的提取,提出了一个新的异形度表征指标CVr2,并在此基础上,将几何矩的思想融入,提出了更为合理的加权改进指标CV(?)。原有的异形度指标体系中,有些不同的截面形状要求用不同的指标,使得形状相差较大的纤维截面形状的异形度指标值,无法直接相互比较。而本文所提出的异形度表征指标CV(?)可以用于表征任意实心异形纤维截面的异形度,无论这个形状是凸还是凹。这使得任意实心截面形状的异形度都可以直接进行相互比较。且与现用的所有异形度指标之间皆无明显相关性,给异形度的表征提供了一种统一的度量标准,真正实现了异形度的定量表征。指标CV(?)的一个优点来自其明显的平移、缩放和旋转不变性。其取值在区间[0,+∞],当且仅当纤维截面为理想圆形时,取最小值0。指标CV(?)的表征能力用一些基本形状和实际异形纤维截面进行了实验验证。结果表明,CV(?)能够有效区分各种纤维截面相对其面积等效圆的差异程度。与对比指标γ(γ=(?))相比,CV(?)在包含深凹部分的形状的异形度表征上更合理。y对于深凹部分所带来的周长大幅增加非常敏感,所得值偏大,而CV(?)没有这个缺点。CV(?)对于CVr2的改进效果,表现在对纤维截面主体外轮廓的波动非常敏感,而对深内凹部分的轮廓波动比较迟钝,在对异形度的表征更符合纤维与纤维外轮廓相接触的实际情况。而且,作为一个基于面积的指标,它是稳定的,抗噪声能力强。在对异形纤维自动识别的研究中,本课题提出了一种基于横截面轮廓波动曲线的特征提取方法和一种基于特征向量之间的非相似度量的顺序聚类识别算法。纤维截面的轮廓波动曲线同时包含了纤维截面形状和大小的信息,能够作为纤维自动识别的有效特征参数。但是,由于实际呈现在图像中的纤维对象分布的随机性和截面变形的不可避免,如何正确地校正纤维对象的起点是一大难题,用图像处理的方法很难解决。本文通过用信号处理技术来计算曲线的互相关,寻找最大值的方法,成功地解决了矫正纤维对象的起点的难题,实现了距离波动曲线的规范化。对于两个异形纤维截面,它们规范化后的轮廓波动曲线的相似程度充分反映了它们本身形状的相似度。因此,本文提出通过提取一个试样片中的所有纤维截面的轮廓波动曲线,两两比较它们的相似程度,自动聚类来识别纤维。本文提出的基于顺序聚类的自动识别算法是基于特征向量的差异性度量的,度量指标采用欧氏距离。该算法的时间复杂度为O(N2)。跟基本的顺序算法方案(BSAS)相比较,时间复杂度稍高,但是它的一大优点在于,对于一个给定的对象,只要它与任一个其所属类的类内对象的距离在阈值线以下,它就能被正确地归类。这样最终聚类的结果并不受送入算法的特征向量的先后次序的影响。用真实异形纤维截面进行了聚类算法的识别验证实验,平均识别正确率达到97%。表明了本课题所定义的基于轮廓波动曲线的形状描绘参数和提出的顺序聚类识别算法能够有效地识别不同截而形状的实心异形纤维。轮廓波动曲线能够充分、有效地表达异形纤维横截面轮廓的特征而用于纤维自动识别,提出的规范化方法也是有效的。其抗噪声能力很强,对纤维的随机分布,图像像素计算的误差,和对纤维在取样、包埋、切片过程中的一定程度的挤压变形不敏感。本系统的结果输出包括两部分：TXT结果信息文档和带编号和类别号的BMP图像文件。其中文档部分是数据的记录,包括所有有效纤维截面对象的面积、长宽比、聚类分类的结果、每一类别形状的具体描述、每一类别组分的根数百分比；图像部分是分类结果直观的图示。在实现了纤维异形度的新指标的表征、计算和横截面形状种类识别这两大功能之后,用各种不同截面形状的异形纤维进行了批量实验,对系统的有效性,稳定性和结果的可再现性进行了分析和测试,并对影响系统识别正确率和测试速率的因素进行了分析讨论。实验结果证明了6种纤维横截面的异形度指标计算的稳定性、6种测试样品的截面形状种类识别准确率的稳定性。系统的再现性测试也取得了良好的效果,对同一试样片进行的异形度计算和自动识别有很高的可再现性。此外,本文还对系统的扩展应用做了进一步的分析。以织物的芯吸性能为例,分析了纤维截面的异形化对纤维集合体的芯吸性能的影响机理,并借助已有的文献资料中的实验数据,对本文提出的异形度指标CV(?)的具体应用做了初步的分析和验证。证明了该指标对织物的芯吸性能具有一定的预测作用,可以用来分析异形纤维,尤其是吸湿排汗类异形纤维的截面异形程度对织物的芯吸性能的影响规律,指出了该指标的具体应用上的一个有价值的研究方向。更多还原

【Abstract】 Fabrics made of profiled fibers with non-circular cross sections can be found on the functional wear market and widely used for sportswear. Changing the cross-sectional shape is the easiest way to alter the meachanical and aesthetic properties of a fiber. This is commonly done by changing the shape of the spinneret hole to produce the fiber shapes desired. Compared to the circular section fibers, fibers with modified cross sections possess more desirable qualities, such as increased luster, firm hand, bulkiness, soil resistance and reduced pilling. As for profiled fiber, quantitative characterization of fiber cross-sectional shape is an extremely important part of component analysis and quality control. It is indispensable when studying the influence mechanism of fiber cross-sectional non-circularity on the performance of end-use products.However, present methods of quantitative analysis on the cross sections of synthetic fibers are usually approached by eye observation and approximate measurement. The current standard test method of synthetic fiber shape factor has its shortcomings. For example, it is quite possible for fibers with different non-circularity degrees but identical measured shape factors. And the profiled fiber identification in cross-sectional view relies on subjective judgments, which is very time-consuming and tiring. Therefore, it is urgent to provide an automated system to complete cross-section shape analysis of profiled fiber.Based on this, the problem fallen into the scope of automatic image analysis system of profiled fiber cross-sectional shape in the condition of established automatic system of acquisiting microscopic images of fibers. An image processing system of profiled fiber cross-sectional photomicrographs is established, and then cross-sectional shape analysis of profiled fibers, quantitative characterization of shape factor and automatic identification of profiled fiber are investigated. We propose a new measure of shape factor which can be used to characterize all kinds of solid cross sections of profiled fibers. We propose an effective method of cross-sectional shape characterization for profiled fiber identification by comparing the extracted distance fluctuation curves of fiber cross-sectional boundary to the centroid. And we complete automatic identification of profiled fiber by means of a new sequential clustering algorithm based on dissimilarity measures among feature vectors.For image processing on profiled fiber photomicrographs, possible types of noises and their distribution are analyzed in this paper. The specificity of microscopic image processing is analyzed from the view of quantitative characterization of shape factor and profiled fiber identification. The pipeline of image processing which is suitable for profiled fiber photomicrographs are presented with two key steps which can greatly improved the accuracy of image segmentation. (1) Based on the characteristics of profiled fiber photomicrographs, an image enhancement technique of maximizing the contrast of the target (the boundaries of fibers) and the background (the texture of fiber cross-section and the body of resin) is proposed; (2) The accuracy of segmentation are greatly improved by computing the smallest convex polygon according to the registered objects. Experimental results show that the noises can be well controlled and the non-uniform illumination phenomenon can be eliminated effectively by the proposed image enhancement algorithm. The method of segmentation post-processing by calculating the smallest convex polygon can well protect the integrity of the boundary of fiber, while easily and accurately removing noises. The accuracy of image segmentation algorithm can be improved significantly. Compared with the common edge detection method, the proposed method is more effective because it can eliminate the double edge and false edge to extract the intact boundary with minimum noises. From the perspective of post-processing, this can reduce the difficulty of feature extraction.For quantitative characterization of shape factor, we present a new measure CVr2 and its modified measure CV(?)(a more reasonable weighted measure based on moments) of profiled fiber shape factor by extracting the distance-versus-angle function of the fiber cross section. Different from the traditional method that based on the inscribed circle and the circumscribed circle, our method is based on geometric moments. Furthermore, CV(?) can be applicable to any arbitrary shape with solid cross section, no matter whether it is convex or concave. Therefore, the values of CV(?) of different cross sections can be directly compared. Since CV(?)does not relevant with any measure of shape factor currently being used, it can provide a unified basis for quantification characterization of fiber shape factor.One of the advantages of CV(?)comes from its obvious translation, scaling, and rotation invariance property. It ranges over the interval [0,+∞) and CV(?)equals to the minimum 0 if and only if the cross- sectional shape is a perfect circle. The behavior of CV(?)is demonstrated on some regular shapes and practical cross-sections. Experimental results show that the defined measure can depict the discrepancy degree between fiber cross-section and the area equivalent circle. Compared withγ(γ=P2/4·π·A-1), CV(?)performs better in the case of cross-sectional shapes with deep concaves,γis sensitive to the change of the perimeter caused by the presence of deep concaves in cross-sectional shapes, whereas CV(?)does not have this disadvantage. Compared with CV,, CV(?)is more sensitive to the convex contours while less insensitive to the contour of deep concaves, more according with the real contacts among fibers. As an area-based measure, it is robust against noise.For automatic identification of profiled fiber in cross-sectional view, we propose an effective method of cross-sectional shape characterization for profiled fiber identification by comparing the extracted distance fluctuation curves of fibers’boundaries to the centroids and we also proposed a new sequential clustering algorithm based on dissimilarity measures among feature vectors.The distance fluctuation curve contains sufficient information on both the shape and the size of the cross section. Therefore, it can be used as the descriptor of cross-sections for profiled fiber identification. However, due to the randomly distributed fibers and shape deformation occurred in the microscopic images, it is hard to caliberate the starting point of fiber object in terms of image processing. In this paper, this challenge is tackled by finding the maximum value on the co-relationship curve, which intrinsically normalized the distance fluctuation curve. For two fiber cross-sections, the similarity degree of their normalized boundary fluctuation curves normalized can effectively reflect the similarity degree of themselves. Based on this, our method extracts the curves of all fiber cross-sections in one slide to compare the similarity degrees between each other, and then creates clusters to identify them. The proposed sequential clustering algorithm is based on dissimilarity measures among feature vectors. The well-know Euclidean distance are adopted as the proximity measure. The time complexity of this algorithm is O(N2).Although it is higher than that of the basic sequential algorithmic scheme (BSAS), for a given object, it can be clustered correctly if the distance between this object and any element of the class is lower than a given threshold. Such clustering result is independent on the order in which the vectors are presented to the algorithm. The experiment of verification of the clustering algorithm using real profiled fiber cross sections has also been done. The average recognition accuracy rate is about 97%. It is shown that the defined shape descriptor and the classification method can effectively identify solid profiled fibers that differ in their cross-sectional shapes. The distance fluctuation curve can characterize profiled fiber cross-sectional contour for profiled fiber identification effectively. The normalization method is also feasible. In addition, this curve is not sensitive to the calculation error caused by the discrete nature of digitial images. It is also insensitive to a certain degree of compressive deformation of cross sections during sample preparation.The resulting output of this system consists of two parts:TXT document and BMP image files. The document records the data of the result of shape factor calculation and fiber identification including area, aspect ratio, tag number, class number and component percentage. The image file is for intuitive output of fiber classification.After achieving the function of quantitative characterization of shape factor and profiled fiber recognition, batch testing is performed to validate the effectiveness, stability and reproducibility of the whole system through various shapes of profiled fiber. And the impact factors of the system are analyzed and discussed. Experimental results show that the shape factor calculation of our system is stable on 6 types of cross-sections and the recognition accuracy is stable too on 6 types of samples. The test on the system reproducibility has achieved good results. The output can be reproduced by using the same slide.In addition, further investigation on the extended application of our system has been made. Taking wickability of fabric as an example, the influence mechanism of fiber cross-sectional modification on the wicking properties of fiber assembles is analyzed. According to the experimental data existed in relevant literatures, we make a preliminary analysis and verification of the usefulness of the shape factor measure CV(?)proposed in this paper on textile industry. It shows that CV(?) has a predictive effect on the wicking property of fabric, and it can be used to analyze the effect of non-circularity degree of profiled fiber, especially for functional fiber with characteristic of moisture absorbing and drying quickly, which implies a potential valuable direction of using our proposed index.更多还原