织物疵点检测的计算机图像分析和评定

【作者】 王朝莉；

【导师】 汪黎明；

【作者基本信息】 青岛大学 ， 纺织工程， 2006， 硕士

【摘要】 本文借助MATLAB编程语言，利用数字图像处理技术和人工神经网络技术对织物疵点的自动识别进行了系统的研究，提出了一种基于图像处理的、以人工神经网络技术来实现分类的织物疵点计算机识别的新方法，并对这一方法的可行性进行了验证。借助于MATLAB的图像处理、小波分析、人工神经网络三大工具箱，通过大量的实验比较和验证，找到了有利于疵点分析的图像处理方法，最优的疵点特征参数的提取，高效而准确的疵点识别模式，将整个图像处理流程系统化，并使其能够有效地兼顾不同种类的疵点图像，形成了一套包括织物疵点图像的预处理、二值分割、分割后处理、疵点部分特征提取、神经网络识别这五个关键环节的、完整的织物疵点自动识别系统。并对其识别能力进行了实验评估，证明了所建立的识别系统，对于简单织物组织的坯布图像，在光影、底色相近的条件下，在经疵、纬疵、区域、破洞四类疵点分类上做到了识别快速、准确。图像处理程序实现过程中，舍弃了获得精确尺寸的意图，着眼于疵点类别的快速、准确分类，做了一些大胆假设。实验证明，所做的假设基本符合实际，可以大大改善识别效果、简化识别算法，且对疵点的个数确定、种类识别和定位无影响。更多还原

【Abstract】 In this thesis, recurring to MATLAB, the detection of fabric defects using digital image processing and neural network technique is studied. A new way of detecting fabric detects which based on image processing and accomplish detect classification by means of neural network is put forward. Furthermore, the feasibility of this way is testified.Recurring to three MATLAB toolbox: image processing toolbox, wavelet toolbox, neural network toolbox, by a mass of experimental comparisons and validations, the image processing methods which applies to detect analysis, the best feature extraction of fabric detects and the efficient and exact detect recognition pattern are found; the systematization of the whole image processing flow is achieved for coping with the different kinds of detects in image processing, accordingly, a integrated system recognizing detects automatically which consists of five key steps: image preprocessing of fabric images with detects, two-value image segmentation, image processing after segmenting, feature extraction of fabric detects and the recognition and classification of fabric detects by means of neural network model is formed. The experimental evaluation of this system is also made and it is testified that this system can quickly and exactly recognize detects in in-gray images which have simple fabric weaves to one of four classes: warp direction detects, weft direction defects, regional defects and discrete defects when the lightness and grounding color are similar.In the process of achieving image processing, the intent of getting the exact sizes of detects is discarded and some important hypotheses are made to classify detects quickly and precisely. And it is testified that these hypotheses according with the fact basically can improve effect and simplify arithmetic greatly and do not counteract classifying , orientating and getting the number of detects.更多还原