【作者】 邹超；

【导师】 汪秉文；

【作者基本信息】 华中科技大学 ， 控制理论与控制工程， 2009， 博士

【摘要】 布匹疵点检测是纺织行业生产和质量管理的重要环节之一，目前人工检测速度慢，劳动强度大，受主观因素的影响，缺乏一致性和可靠性。计算机视觉技术随着大规模集成电路和图像处理技术的发展，在工业表面检测领域有了越来越多的应用，基于机器视觉的检测系统已成为布匹检测未来发展的必然趋势。文中首先指出单一的方法并不能在实时条件下完成疵点检测所有任务，并首次提出了疵点在线检测系统的三层结构，分别完成疵点判别、疵点分割和疵点分类三个功能，并设计了检测系统与之对应的各部分硬件和软件结构，三层结构大大提高了系统实时性。文中分析造成图像灰度不一致的两个主要因素后，分别提出了光照空间分布不一致和光源波动的在线校正算法。本文提出的校正方法均不需要独立标定，容易实现、准确率高，能大大改善了图像质量。本文提出了模糊类别共生矩阵及其特征和基于距离的异常点检测方法作为疵点判别的实时算法，这种方法中色调集合替代了灰度级。文中论述了色调集合的划分、模糊色调集合隶属度函数的拟合和间距参数的选取等问题，并讨论了该算法的实时实现。实践证明该方法在保证了检测准确率的同时，实时性能上远远优于已提出的多数方法。本文通过定义分辨力函数，以合成疵点图像为研究对象，提出了在实时应用中对各种不同的纹理图像能快速确定Gabor滤波器组参数的方法：疵点分割中只需用到实Gabor滤波器；方位角只需取水平和垂直方向，径向中心频率和FIR滤波器长度的选取依赖于布面纹理的固有频率。本文提出的方法使得滤波器个数减少到4个，大大提高了算法的实时性的同时，仍能正确地分割出大多数的疵点。且整个疵点分割过程无需离线标定参数，对噪声不敏感。本文分别采用基于Blob分析和基于类别行程共生矩阵来恢复疵点全局信息，并提取疵点区域的几何特征和异常行程长度作为疵点评分依据。本文提出的两种方法相比许多监督分类方法而言，不需要事先收集到疵点样本，也不需要离线标定参数，就能成功地将疵点分为三大类进行评分。本文还探讨了基于蚁群聚类算法的疵点检测方法，该算法固有的并行分布式计算结构是提高系统实时处理能力的重要手段。受到蚁群聚类算法的启发，疵点可看作是由一些局部不规则点的集合在空间上聚集而成的纹理图像部分，本文详细探讨了蚁群算法的三大要素：蚂蚁环境度量，即局部不一致特征定义，信息素释放与挥发机制和蚂蚁决策。进一步地本文将每只蚂蚁都当作一个独立的模糊推理单元，蚂蚁的环境度量用一组语言变量来描述，并通过模糊推理规则，确定蚂蚁的移动速度和移动方向。更多还原

【Abstract】 Fabric defect detection is one of most significant procedure for quality control in textile manufacturing industry. The manual detection is time-consuming, labor-intensive and devoid of consistency and reliability due to many subjective factors. As a analog of human vision, computer vision technology has been applied widely in industrial surface detection with the advances of digital integration and digital image processing techniques. It can be confirmed that the computer vision based detection system has bright perspective of the automation of fabric defection detection.In this dissertation, initially it is pointed out that the overall task of defect detection can not be accomplished by any single method, and the detection system is divided hierarchically into three levels containing defect judgment, defect segmentation and defect classification. The designs of the hardware and software supporting the corresponding level are discussed respectively. It is also analyzed that the inconsistency in the captured images can be attributed to the spatial variation and noise of the illumination source. The online algorithms are proposed to rectify both inconsistencies to improve the image quality. These algorithms are accurate and facile to be implemented without any additional calibration procedure.The Fuzzy Label Co-occurrence Matrix with its features and the distance based outlier detection are proposed as the real-time defect judgment algorithm, which is based on the tonal classes’ substitution for gray levels. The texture tone classification, the fitting for membership functions of fuzzy tonal sets and the spacing parameters selection are detailed. The real-time implementation of the algorithm is also discussed. The proposed method has much greater real-time performance than other methods while maintaining high accuracy.By defining a discriminability function and utilizing some images with synthetic defects, a parameter selection method is proposed to adapt to various fabric types when the Gabor filter bank is applied in defect segmentation. It is argued that only the real part of Gabor filter works during segmentation; the radial orientation can be limited to horizontal and vertical direction; the radial frequency and the length of the FIR filter should depend on the basic frequency of the fabric texture. Though the number of the filters decreases to four for better real-time performance, the Gabor filter bank can segment most defects accurately. Moreover, the whole defect segmentation is independent of an offline calibration and insensitive to noise.The blob analysis based method and the Label-Run Co-occurrence Matrix features are applied to attain the overall area of each defect. The defect is therefore classified into three categories with a certain remark based on the geometric features of the blob area or the exceptional run lengths. Compared to methods based on supervised classification, the proposed methods excel in that they don’t need any pre-collected defect samples or any offline calibrations.Finally, the ant colony clustering algorithm is introduced into the area of defect detection since the intrinsic parallel distributive computation architecture values in the real-time performance of the system. Inspired by ant colony clustering algorithm, defects are considered to be the gathering of the locally irregular pixels. Three basic elements of ant clustering algorithm are detailed: ants’ environment measurement, that is, local irregularity measurement, pheromone release, diffusion and evaporation and the behavior decision of ants are dissertated in details. Further, the ants are considered as fuzzy inference units and the local irregularity features are then transformed to be depicted by some predefined membership functions, and the speed and direction of ants are inferred by fuzzy rules. Finally the defect detection algorithm based on fuzzy ant colony clustering is consequently proposed.更多还原