【作者】 宗亚宁；

【导师】 王善元；

【作者基本信息】 东华大学 ， 纺织材料与纺织品设计， 2006， 博士

【摘要】 本文以棉麻纤维图像特征提取和识别为研究对象，系统研究纺织纤维自动检测的技术，主要包括纤维切片技术、图像自动摄取技术、图像轮廓描述技术、图像分割技术、图像特征提取和纤维识别技术。纤维成分的测定在纺织加工业、商品交易和国际贸易中有极其重要的地位。对由化学结构相似纤维制成的混纺产品，无法用化学溶解法测定其成分比例，通常只能使用显微镜鉴别方法，棉／麻纤维的鉴别是其中的典型代表。长期以来，人工检测为基础的检验方法，检验时间较长，检验成本较高，检验结果会受主观情绪和长时间工作的视觉疲劳影响产生测试误差。本文研究棉／麻自动检测的相关技术，从而达到缩短检验周期，提高检验效率，减轻检验人员的劳动强度，排除人为因素的干扰，保证结果的一致性和客观性。纤维切片技术关系到纤维自动检测的难易和检测结果的准确性。本文首先简单介绍制备纵向形态测试样品的过程，利用煮沸法形成分布均匀、密度适当的纤维悬浊液，然后在载玻片上形成分布均匀的纤维层，并且尽可能减少载玻片上纤维之间重叠和交叉。根据纤维图像自动检测和分析的需要，采用预聚技术的丙稀酸类树脂快速包埋技术制作纤维截面切片，改进重点在于降低该类包埋剂聚合时收缩率，避免不均匀聚合，同时提高包埋速度，以适应常规研究和快速检测的要求。通过对木棉纤维和涤纶单中空纤维中空率、圆整度进行测试，证明本技术具有良好的保形性能。采用树脂切片技术获得的纤维截面切片，切片薄，截面变形小，截面之间的粘连少，为图像自动检测系统的研制打下良好的基础。图像自动摄取技术保证图像数字化的自动进行，该技术主要由三部分组成——自动聚焦、多焦面图像融合和图像拼接技术，核心是聚焦算法。在聚焦算法研究上，对灰度差分算子、灰度方差算子、Tengrad算子、Lapal算子和灰度熵算子在纤维显微图像聚焦效果进行研究，在无偏性、单值性、单义性、抗干扰性和简单性五个方面进行性能比较，依据聚焦曲线选定灰度差分算子作为纤维图像粗聚焦算子；通过对纤维图像的特点进行分析，提出自适应局部区域和基于轮廓边缘灰度变化精确聚焦方法，分别应用于纵向片段和截面聚焦。通过对实际摄取过程的分析，充分利用不同视野聚焦平面位置之间的相关性，预置起始聚焦位置，缩短聚焦距离，和改进的调焦策略相结合，提高聚焦速度。利用自动聚焦过程提供的多焦面图像聚焦信息，在粗聚焦图像的基础上，根据纤维轮廓，将多焦面图像组中每幅图像分割为背景区域(无目标)、清晰区域(目标聚焦区域)、模糊区域(目标离焦区域)。融合图像背景由系列聚焦图像背景区域平均获得，目标区域通过清晰区域直接替代模糊区域获得，并结合边界处理。在显微图像处理分析中，受显微镜视野的限制，常常需要将多幅图像拼成一幅较完整的图像。本文提出的拼接算法基于模板匹配原则，根据图像的相邻关系确定拼接模板和拼接区域，利用中值滤波和梯度算子处理拼接模板和拼接区域，避免噪声干扰和图像灰度差异影响，保证定位精确和拼接正确。目标轮廓的特点是纤维图像识别研究的重点，利用链码对目标轮廓描述，可以将轮廓的2-D边界形状描述问题，转化成对1-D波形进行分析的问题，为随后进行的目标形状特征的描述和判别提供便利。为解决图像噪声干扰问题，采用轮廓边缘平滑方法；为提高链码表征角度的能力，扩大计算轮廓切线方向的支撑区间，提高抗噪声能力和减少线性采样的影响，对Freeman链码进行改进，引入方向链码概念，即累加差分链码和；通过对人工轮廓方向链码曲线的分析，进行轮廓角点提取技术研究。针对纤维纵向形态图像，利用方向链码确定轮廓跟踪起始点和判断纤维交叉部位及纤维头端；根据纤维纵向轮廓的特点，提出快速生成纤维中轴线的算法。将纤维截面粘连形式分为串联、并联两种基本形式，给出判断截面粘连的算法，针对不同的截面粘连形式提出不同的分割策略。最后对典型棉麻截面的中腔胞壁轮廓进行方向链码描述，并归纳了它们的链码曲线特点。将轮廓采用链码表示就意味着将二维的图像分析问题转化为一维信号处理，就可以利用信号分析技术来提取纤维截面特征。从信号分析角度，轮廓信号可视为由轮廓包络信号和轮廓特征信号组成。小波变换通过伸缩和平移等运算功能对函数或信号进行多尺度分析，具有多分辨率分析的特点，很适合探测正常信号中夹带的瞬态反常现象并展示其成分。文中首先对小波变换原理、多分辨率分析原理及一维Mallat快速算法进行简单介绍；对纤维截面轮廓的方向链码曲线振荡原因进行分析，依据小波降噪的原理，利用多层小波变换进行信号降噪；利用小波变换多尺度分析能力，先获得截面轮廓的包络信号，然后获得轮廓特征信号，并在裂缝、凹陷和角点的模型上验证本算法的可行性；最后对中腔胞壁的突变特征信号进行相关性分析，利用相关性系数区分棉麻纤维截面。在棉麻纤维纵向形态识别方面，首先简单回顾已有的棉麻纤维纵向形态特征参数；根据纤维纵向轮廓的特点，提出快速准确计算纤维纵向形态投影直径的算法；根据纤维纵向形态图像灰度变化的特点，提出利用纤维断面灰度变化特征表征棉转曲和判断棉中腔的方法；针对纤维横节和裂纹提取的困难，采用灰度投影方法提取纤维横节和裂纹特征，将二维纹理特征的提取转化为对一维灰度曲线波形分析；引入图像样式转换技术解决弯曲纤维横节和裂纹提取困难，并在此基础上提出基于纤维中轴线的快速判断方法；最后建立棉麻纤维纵向形态识别流程。针对天然纤维特征参数不稳定的特点，利用神经网络进行纤维识别。首先对棉麻纤维截面特征参数的分布进行统计，初步得到它们的分布规律；然后在介绍神经网络原理和特点的基础上，构建用于纤维识别的三层BP神经网络；采用固定测试法和交叉检验法对特征参数在识别中的作用进行了研究，筛选出用于识别的截面特征集合；并在此基础上，对网络隐层节点数目和激活函数类型进行优选。针对天然纤维特征参数不稳定的特点，利用神经网络进行纤维识别。首先对棉麻纤维截面特征参数的分布进行统计，初步得到它们的分布规律；然后在介绍神经网络原理和特点的基础上，构建用于纤维识别的三层BP神经网络；采用固定测试法和交叉检验法对特征参数在识别中的作用进行了研究，筛选出用于识别的截面特征集合；并在此基础上，对网络隐层节点数目和激活函数类型进行优选。本文以棉麻纤维为研究对象，系统研究纺织纤维自动检测的相关技术，获得了一系列成果。首先改进树脂包埋技术，缩短包埋时间，为纤维截面分析提供低变形、高分离度的光学显微镜半薄切片，保证截面图像质量，基本满足实际检测工作的需要；根据纺织纤维光学显微图像的自身特点，确定自动聚焦算法，为自动显微镜的研制和应用奠定基础；对Freeman链码进行改进，引入方向链码概念，可以更精确地描述纤维截面的轮廓特征，而且将二维图像特征提取转化为一维信号分析，方便随后的截面特征分析工作；利用小波变换具有多尺度分析的特点，对轮廓细节特征进行提取，准确反映截面轮廓特征变化；充分利用纤维图像中的灰度信息，增加分析信息来源，提高纤维识别率；最后将神经网络技术应用到纤维判别，利用其自学习、自组织的特性应对天然纤维截面特征参数分布波动的问题。上述技术不仅仅只适用于纤维成分分析工作，还可以推广到纤维品质检测、纤维成形技术、纺纱工艺过程质量控制以及纺织材料结构检测等领域，具有广阔的应用前景。更多还原

【Abstract】 The feature extraction and determination of cotton and bast fiber is researched in this thesis. A series of automated determining technology in textile fibers is studied, which include microscopical sample preparing technology, image automatic photographing technology, the object contour describing technology, the image segmented technology, feature extraction technology and fiber identification technology.The determination of fiber contents is very important in textile industry, national and international trade. The chemical dissolving analysis is no use for these textiles containing two or more fiber types of similar chemical structure. The only way is microscopical analysis, and the determination of fiber content in cotton/blast fiber blends is typical in this identification type.Microscopical analysis is based on manual determining, which results in much time-consuming, high testing-cost. The identification result can be easily influenced by subjective emotion and eye strain for long time work. The automatic determination of fiber contents can shorten inspection period, heighten testing efficiency, lighten technical labor intension, eliminate artificial disturb, ensure results consistency and objectivity.The slicing technology of fiber directly influences the difficulty and correctness of automatic identification. The procedure of preparing the longitudinal slices is simply introduced. A uniform and fairly fiber suspension is obtained through quick boiling, the noticing part of which is to avoid the fiber overlapping and crossing. The improved methacrylate resin embedding is used in fiber cross-section slices. This method reduces the polymerized shrinkage, uneven polymerization and embedding time. The embedding method is fit for research and fast testing. The hollowness and circularity measuring from the slices of hollow polyester fiber and kapok proved this method has best form-keeping. The slice from resin embedding has little deformation, low touching and thin thickness. These will facilitate the automated image testing system. The automatic photograph ensures the automated image-digitalizing. This technology is made up of auto-focus technology, multi-focus image fusion and image stitching, the core of which is auto focusing algorithm. In the part of auto-focusing, several focusing algorithms are discussed: the sum of the absolute of value of gray difference(SMD), the gray scale variance, the Tengrad operator, the Lapal operator and the gray scale entropy. The unbiased ness, uniqueness, single meanings, anti-noise performance and simplicity of these algorithms is compared, and the SMD is selected for coarse focusing algorithm based on the focusing curve. The adaptive local sub-area algorithm is applied in the fiber longitudinal section image, and the gray variation around the contour is applied in the cross sectional focusing. The practical focusing procedure is analyzed, and the relativity between locations of focal planes of different visual fields is used to decide the pre-location of focusing procedure. The pre-location method which shortens the focus distance and the improved focusing tactic increase focusing speed.At the use of the info of multi-focus image from auto-focusing procedure and coarse focusing image, the images from the image series are segmented into background(no object), sharp area(the object focusing) and blur area(the object defocusing). The background of the fusion image is obtained from the average of backgrounds of series images, and the object area of the fusion image is obtained through the sharp area substituting blur area and combining with border processing.In microscopically image analysis, several images is stitched together to form an entirely image for the limitation of visual field. The algorithm is based on template matching principle. The stitched template and area is defined from the relationship between neighboring images. The two areas are processed by median filter and gradient operator in order to escape from the influence of noise and image gray level difference, which ensures the correctness of location and join.The character of object contour is the focus of fiber image recognizing. The feature extraction of 2-D image is converted into the analysis of 1-D signal by the use of chain code describing method for the contour, which facilitates the following feature analysis of shape. The border smoothing is used to decrease the influence of image noise. The concept of direction chain code is introduced, which is the sum of differential chain code cumulativeness to increase the ability of direction expression, anti-noise and anti-rectilinear-sampling. Depended on the direction chain code curve from artificial contour, the method of detecting corner is studied. Aiming at longitudinal image, the start point of border-tracing, the ends and cross part of fibers is decided on the directional chain code. The middle axis can be quickly decided on the character of longitudinal shape. The touching cross-section can be divided into two basic touching type, serial touching and shunt touching. The algorithm to judge touching cross-section and the different separating algorithm for different touching types is brought forward. The lumen and cross-section border of the typical fiber/ramie fiber is described by the direction chain code, and the character of these curves is concluded.If the contour is described by the chain code, this means that the 2-D image analysis can be converted into 1-D signal proceeding, so the signal analyzing technique can be used to extract features of fiber contour. The contour signal can be taken for the buildup of envelop signal and detail signal. The wavelet transform have multi-scale analysis on signal through dilation and translation transform, which is suitable for detecting the transient abnormal phenomena and its component. The principle of wavelet transform, multi-resolution analysis and 1 -D Mallat fast method is simply introduced. The reason for the surge of directional chain code curve is analyzed, and the curve is denoised by the multi-level wavelet transform. The envelop signal is obtained from the contour curve by the multi-scale analysis, and then the detail signal is gotten. This method is validated on the crack model, concavity model and corner model. The correlativity between lumen and cross-sectional border is calculated, and the coefficient is used to discriminated cotton/bast fibers.In the part of distinguishment of longitudinal modality, the distinguishing algorithms which have been studied are simply introduced. A fast algorithm of measuring fiber widths is put forward based on the fiber longitudinal character. The gray level variation is used to decide natural convolution and lumen of cotton. The gray projecting method is used to extract the cross markings and cracks, which converts 2-D feature extraction into the 1-D gray curve analysis. The difficulty of extraction of the cross markings and cracks on curly fiber is solved by the image transform, and a fast improved deciding method based on fiber middle axis is put forward. The flow of deciding fiber type on the longitudinal feature is built up.The neutral network is used to determine the fiber types to answer for the unsteadiness of natural fibers feature. The distributing of cross-sectional character parameter is investigated, and the elementary statistics rule is obtained. After the principle and trait of neutral network is introduced, the three-level BP neutral network is formed for fiber determining. The effect of feature in fiber determining is researched on steady testing method and cross testing method, from which the determine set of cross-sectional feature is screened out. Based on these work, the hide-layer nodes number and activating function is choose.This thesis focuses on the determination of cotton/blast fiber. The determining technologies are systemically researched. First the embedding resin is improved, and the embedding time is shortened. The little deforming and high separating half-thin microscopy slide is provided, which ensures the cross-sectional image quality and satisfies the practical testing task. The auto-focusing algorithm is decided by the character of microscopical image of textile fiber, which is the foundation of auto-microscope. The Freeman chain code is improved and the concept of direction chain code is introduced. The new code describes the contour of fiber cross-sectional more precisely. The feature extraction of 2-D image is converted into the analysis of 1-D signal, which facilitate the following feature analysis of shape. The detail feature is extracted from the multi-resolution analysis of wavelet transform, which indicated the character more truly. The gray level info is fully made use of to increase the analysis source and the determining correctness. The cross-sectional image is differentiated by the artificial neutral network, of which the self-study and self-organize is fit for the unsteadiness of the natural fiber cross-sectional feature. All above technology is not just fit for fiber content analyzing, but also can be used in fiber quality inspect, fiber forming technology, the quality control of spinning procedure and the testing of textile material structure, which has extensively applied fields.更多还原