【作者】 景军锋；

【导师】 贾建援；

【作者基本信息】 西安电子科技大学 ， 机械电子工程， 2013， 博士

【摘要】 为了提高纺织工业产品的可靠性以及重现性，对织物外观特征进行客观评价起着越来越重要的作用，比如疵点、起球、组织结构、色差以及粗糙度等等。本文采用数字图像处理技术开展对织物疵点自动检测和分类、织物起球客观评级体系以及织物组织结构识别及分类等三个方面进行了基础理论的研究，所取得的主要研究成果为：1.提出一种基于改进的Gabor滤波方法、数学形态学处理法和多尺度小波检测的方法库的系统检测法.首先采用改进的Gabor滤波方法，选出最优滤波结果，进行高斯平滑，确定出正常织物图像的两个阈值门限，进而分割出织物的疵点图像；其次采用形态处理学的处理方法对织物图片进行检测；最后采用了多尺度小波检测的方法，检测最终结果。2.提出一种基于局部二进制模式与Tamura纹理特征方法相结合的织物疵点分类算法。该算法主要完成的任务是对织物特征向量的提取，局部二进制模式从局部或像素邻域描述纹理的特征，Tamura纹理特征方法从全局描述疵点纹理特征，两者结合能更好的描述疵点纹理特征。完成特征向量提取后，选用共轭梯度BP算法来处理特征向量。共轭梯度BP算法收敛性较好，提高了训练速度和训练精度。实验结果表明，提出的算法对疵点较高的分类准确率。3.印花织物分为周期印花织物和无规则印花织物，在周期印花织物检测中，基于织物图案信息的周期性，规则带通过设定移动窗口，计算移动窗口的能量和方差，以无疵点的织物图像的能量和方差作为阈值，由计算的能量和方差值来检测疵点。在无规则印花织物检测中，通过遗传算法选择Gabor滤波器的参数，最优参数构造的Gabor滤波器匹配最佳检测织物纹理，从而提取织物的有效信息，达到分割织物疵点的目标。4.基于二维离散小波变换和局部二值模式的织物起球客观评价方法。起球特征向量是由提取小波分解后尺度4-6的细节子图像在三个方向（水平、垂直、对角）上的小波能量值和将小波重构尺度3-6细节子图像的重构图进行LBP变换提取其特征值组成。特征值经过归一化处理后通过主成分分析（PCA）来降维，处理后的特征值作为支持向量机（SVM）的数据输入为织物起球等级分类。5.提出了一种自动及实时的分类方法分析三种机织物即平纹，斜纹和缎纹组织。首先，用灰度共生矩阵和Gabor小波方法提取特征值。用主成分分析方法处理纹理特征向量获得最小冗余及最大主成分的特征向量。在分类过程中，应用概率神经网络分类三种基本机织物。实验结果表明有快速的训练速度的概率神经网络分类器能够精确及有效的分类机织物。比较灰度共生矩阵和Gabor小波两种方法，融合两个特征向量获得了最好的分类结果（95%）。更多还原

【Abstract】 In order to improve the reliability and reproducibility of textile industry products,evaluating appearance characteristics objectively on fabrics plays a more and moreimportant role in textile industry, such as characteristics of defect, pilling, organizationstructure, color and roughness, et al. Processing technology of digital image is appliedto three aspects on the basis of theory research including automatic defect detection andclassification of woven material, objective rating system on fabric pilling, fabricstructure recognition and classification in this article. Main research results are obtainedas follows:1. Joint fabric defect detection method based on method library. A joint fabricdefect detection method based on "method library" is proposed, and various effectivedefect detection algorithms on fabrics are improved and synthesized in joint fabricdefect detection method based on method library. The proposed method integratesGabor-Gauss method, background analysis, multi-scale wavelet method as joint defectdetection method library to detect defects on fabrics. A preliminary graphical userinterface (GUI) based on method library is designed in order to facilitatehuman-computer interactions.2. Fabric defect classification algorithm based on combining texture feature oflocal binary pattern (LBP) and Tamura. Fabric defect classification algorithm viacombining texture feature of local binary pattern (LBP) and Tamura texture feature isput forward due to a single feature can not be effective description of fabric defect.Local binary patterns can describe characteristics of fabric texture from local, Globaldefect texture feature could be represented by way of Tamura method. Better descriptionof defect texture feature could be obtained by means of combining local binary pattern(LBP) and Tamura. Conjugate BP is applied to train and test extracted feature vector inproposed method.3. Defect detection on printing fabrics. The printing fabric can be divided intocycle printing fabric and random printing fabric. In the test of cycle printing fabrics,moving window is set in regular band based on the information of periodic patternfabric. Energy and variance of defect-free fabrics could be determined as threshold.Calculation of energy and variance of sample cycle printing fabrics could realize defectdetection based on obtained threshold. In the testing of random printing fabrics,parameters of Gabor filter can be selected via genetic algorithm, Gabor filter structuredby optimal parameters is matched with fabric texture so as to extract the effective information of woven fabric in fabric defect detection, thus the purpose of fabric defectsegmentation can be achieved.4. A new method based on wavelet transform and local binary pattern (LBP) wasproposed for fabric pilling objective evaluation. The surface of the fabric pillinginformation was inspected by using two-dimensional discrete wavelet transform(2DDWT). The pilling feature vectors were consisted of the wavelet energy value whichis the wavelet decomposition of sub-image details in scale4to6with three directions(horizontally, vertically, diagonally) and the LBP features which is waveletreconstruction image in scale3to6using LBP. It is necessary to normalization thefeatures using principal component analysis (PCA) to reduce the dimensions. Then, theprocessed feature can act as fabric pilling classification data input for support vectormachine (SVM).5. An automatic and real-time classification method is proposed to analyze threewoven fabrics such as plain, twill, and satin weave. The methodology involves twoapproaches to extract texture features using gray-level co-occurrence matrix (GLCM)and Gabor wavelet. Then, principal component analysis (PCA) is utilized to deal withthe texture feature vectors to gain minimize redundancy and maximize principalcomponent feature vectors. Finally, for the classification phase, probabilistic neuralnetwork (PNN) is applied to classify three basic woven fabrics. With strong real-time,robustness, fault-tolerance and non-linear classification capability, PNN can be apromising tool for classification of woven fabrics. The experimental results show thatPNN classifier with faster training speed can classify woven fabrics accurately andefficiently. Besides, compared with GLCM method and Gabor wavelet method, thefusion of the two feature vectors obtain the best classification result (95%).更多还原