【作者】 张扬；

【导师】 蒋高明；

【作者基本信息】 江南大学 ， 纺织科学与工程， 2014， 博士

【摘要】 贾卡经编针织物由于贾卡导纱针的不同偏移规律可形成“厚”、“薄”和“网孔”等垫纱效应，再配以其它工艺参数的多元变化，形成独特的多层次视觉效应，精致而丰富的花纹图案，以及柔软舒适的质感，而深受市场和消费者喜爱，产品适用于高档内衣、蕾丝辅料以及室内软装饰等纺织品中。在贾卡织物设计过程中，尽管目前经编CAD系统的二维绘图和工艺处理功能已基本完善，但对于贾卡织物图像的花纹分割等前期处理功能，还需借助于第三方图像处理软件的套索和魔术棒等工具近乎手工地完成织物图像的分割和提取工作。该过程十分耗时且繁琐，将近占用了整个设计过程的70%时间。可见如何改变这种耗时费力的设计方式，快速、准确地获得贾卡经编针织物的花纹图案是经编CAD系统一大亟待于解决的问题。论文以贾卡经编针织物为研究对象，采用计算机数字图像处理技术，对织物图像的纹理特征提取和分割方法展开了系统研究。各章节主要内容概述如下：第一章介绍了论文研究目的和意义，详细概述了国内外文献研究现状和现有纺织CAD系统中织物图像分割技术，分析了贾卡织物图像花纹分割难点所在，并提出了论文主要研究内容和创新点。第二章阐述了贾卡经编机机构组成、起花原理以及贾卡经编针织物分类和特点，分析了贾卡织物图像纹理和噪音信息形成的原因，并着重研究了高斯滤波和双边滤波算法基本原理、不同的参数配置对贾卡织物图像影响。其中，双边滤波采用动态加权系数，在削弱噪音信息，平滑织物图像的同时，也保护了边缘细节，更适合于贾卡织物图像的预处理。第三章实现了经预处理后的贾卡织物图像的多分辨率尺度分解。首先介绍了小波变换基础理论，探讨了贾卡织物图像小波金字塔式结构和小波树型结构两种分解模式，并提出了一种基于图像能量等纹理特征的分解模式判别法则，使贾卡织物图像纹理特征得以充分利用。通过多分辨率尺度分解可以简化贾卡织物图像模型，降低计算工作量，同时为后续的分割模型提供多层次的细节特征，尤其是局部细节特征。第四章主要内容为基于小波域的贾卡织物图像改进K均值聚类花纹分割方法。首先概述了传统图像分割方法和K均值聚类基本工作原理，并针对传统K均值聚类算法在贾卡织物图像分割过程中存在的问题，例如随机选取初始聚类中心，对贾卡织物图像中的大量噪声信息过于敏感等问题提出改进。在改进的K均值算法中，采用小波多分辨率分解对噪声信息去相关性，密度函数和数据邻域等算法优化初始中心点的选择方法，并根据各尺度上频带特征矢量的离散程度给予不同的权重值，增强或削弱特征分量在K均值聚类过程中的作用。本章最后进行算法的比较实验，结果证明该分割算法与改进前相比，准确率更高，适宜花纹数量较少，质地纹理较为细腻的贾卡经编针织物图像。改进的K均值聚类算法除了作为一种单独的分割算法外，同时能在后续马尔可夫随机场（Markov Random Field，MRF）模型中的低分辨率尺度的花纹分割中发挥作用。第五章主要内容为贾卡织物图像多分辨率MRF建模与花纹分割。首先阐述了传统MRF模型及其在贾卡织物图像分割中的应用。再者，针对传统MRF模型中存在的势函数取值过于依赖人工经验，特征场模型中对于图像中的噪音信号问题考虑不足，以及期望最大算法的迭代终止条件等问题，论文提出一种自适应权重的贾卡织物图像多分辨率建模与分割算法，其中通过一自适应权重函数来调整特征场模型和标号场模型在分割过程中的控制权比重，随着分辨率的增大，权重函数取值逐渐减小，分割过程的主导地位由特征场能量控制转向标号场能量控制，分割结果逐步得到精细化，同时可以明显削弱依据经验给定的势函数取值对分割结果的影响。同时，面对多尺度MRF建模过程中，常采用空域非因果方式建模，以及计算过程负载较大难以达到全局优化等问题，提出一种基于MRF层次模型的贾卡织物图像花纹分割算法，算法中采用融合伽马分布和高斯分布的有限通用混合模型来逼近图像的小波系数，零均值隐状态的混合概率分布来描述织物的噪音信息，标号场模型融合尺度间的因果方式和尺度内的非因果方式建模，并采用SMAP参数估计准则通过非迭代方法得到在最高分辨率尺度上的分割结果。本章最后通过与不同算法比较实验，结果表明该算法计算速度和分割准确率均能达到一定的设计要求，适用面较广。第六章对全文做了总结和展望。给出了论文主要结论的同时，提出了进一步完善研究的设想。更多还原

【Abstract】 Due to the factors such as weave structure, fiber and dyeing-and-finishing, jacquardwarp-knitted fabric is multifarious in patterns and exquisite in workmanship, for which thejacquard fabric is very popular and widely used in high-grade lingerie, garment accessory andfurnishing fabric. Although the function of2D drawing and process configuration in currentCAD system of warp knitting is nearly perfect, the pretreatment such as pattern separation ofjacquard fabric is done by using simple mapping tool such as lasso and magic wand. So it isan extremely repetitious, laborious and time-consuming work ranging from a few hours toseveral days, which occupies too much time and will increase production costs. In this case, todevelop a rapid, efficient and automatic pattern separation system for jacquard warp-knittedfabric is rather urgent.This paper focuses on texture characteristics and pattern separation methods of jacquardwarp-knitted fabric by means of computer image processing. The content of each chapter isbriefly introduced as follows.In Chapter1, the research purposes and significance are introduced briefly. Overseas anddomestic research statuses, the current pattern segmentation methods in warp knitting CADsystem are summarized. Then the difficulty in pattern segmentation is analyzed. The researchtopics and innovative points of this paper are proposed.In Chapter2, the machine structure, weaving mechanism and the classification ofjacquard warp-knitted fabrics are introduced. Then the underlying reason for fabric textureand noise signals of jacquard warp-knitted fabric are analyzed. Finally, an algorithm whichcan smoothen the fabric image, weaken noise signal and protect the detail information ofmarginal region is proposed.In Chapter3, the multi-resolution wavelet decomposition of the pretreated jacquardwarp-knitted fabric image is described. Firstly, the background of wavelet transform and thetwo decomposition models such as pyramid structure and tree structure are introduced briefly.Then a discriminant rule of decomposition model for jacquard warp-knitted fabric is proposed.The analysis result indicates that multi-resolution wavelet decomposition can simplify themodel jacquard fabric, lessen the computational burden, and provide the multi-level detailcharacteristic, especially the local detail characteristic.In Chapter4, this paper focuses on the modified K-means clustering algorithm inwavelet transform for jacquard warp-knitted fabric. Firstly, the traditional image segmentationmethods and the mechanism of K-means clustering are introduced. The problems oftraditional K-means clustering for jacquard warp-knitted fabric are analyzed, such as randomchoice of initial clustering center and the susceptivity to noise information of jacquard fabric.Then a modified K-means clustering algorithm is proposed, which includes waveletmulti-resolution decomposition, the optimized initial clustering center and weighting factorbased on dispersion degree. The modified K-means clustering algorithm is not only aindependent algorithm, but also plays an important role in successive chapter.In Chapter5, the concentrates on multi-resolution MRF model for pattern separation of jacquard warp-knitted fabric. Firstly, the traditional MRF model is introduced briefly.Secondly, on account of the problems of potential function, which relays too much onartificial expertise, and feature field which takes insufficient account of noise signals, thepaper proposes a multiresolution Markov random field with adaptive weighting in waveletdomain. The proposed algorithm can control the ratio of feature model energy to label modelenergy using a adaptive weighting function. Thirdly, by reason of non-causal label model andthe computation burden of iterations, the paper proposes a new pattern segmentation based onhierarchical Markov random field model. In new algorithm the label model takes into accountof not only the relationship between global and local, but also the modeling methods such asinter-scale causal model and intra-scale uncausal model. Finally, the segmentation results areobtained by SMAP parameter estimation which is a un-iterative algorithm in originalresolution scale.In Chapter6, the summary is introduced, which includes the main contributions and theproblems of the present study.更多还原