【作者】 张健；

【导师】 鲍际平； 韩宁；

【作者基本信息】 北京林业大学 ， 森林工程， 2010， 博士

【摘要】 强化木地板符合国家可持续发展的资源政策,在地板行业发展迅速,未来发展空间巨大。强化木地板的外观质量缺陷是影响其使用性能的一项综合性指标。目前,国内仍然依靠人为目测对强化木地板外观质量进行检测,对此,本文提出了基于机器视觉的强化木地板表面质量检测方法的研究课题。本文选用原北京柯诺森华地板厂生产的特定系列产品作为研究对象,研究缺陷包括生产中常见的干花、缺纸、污斑和纸裂四种。基于机器视觉的强化木地板表面质量检测主要包括强化木地板的图像获取、图像分割、特征参数的计算与提取和分类器设计几部分。图像分割步骤需要完成把缺陷图像从复杂的背景中提取出来的任务;在特征参数的计算与提取环节,需要在分析地板表面图像特点的基础上,选取能够代表合格地板或缺陷地板特征的参数,同时,为使提取参数达到具有可区别性、可靠性、独立性和数量少的要求,降低后期分类器的复杂程度和运算耗时,需要对特征参数进行降维;在分类器设计阶段,需要设计一种分类器能够将合格的强化木地板和存在缺陷的强化木地板赋以不同的表现形式加以识别。本研究针对强化木地板表面质量检测方法进行了比较系统的研究,主要工作和结论有以下几点:1、在图像分割研究中,论文采用最大类间方差法、基于二维空间的蚁群算法和基于最大熵的遗传算法三种方法对强化木地板图像进行图像分割处理,比较了三种方法针对强化木地板表面缺陷检测的适用性。研究结果显示,最大类间方差法对浅色干花缺陷分割效果不理想;而基于二维空间的蚁群算法对深色缺纸、污斑和纸裂分割效果不理想;基于最大熵的遗传算法对浅色和深色的缺陷分割效果都比较理想,适用于强化木地板表面质量检测的图像分割处理。2、在特征提取的研究中,本文在分析强化木地板表面图像特点的基础上,提出通过计算地板图像的颜色特征和纹理特征来表达地板图像特性。在颜色特征参数计算中,研究将HIS三维颜色空间通过加权求和降至一维空间,在该一维空间中计算颜色的一阶矩、二阶矩和三阶矩参数。由于三阶矩中出现了复数情况,研究将实部和虚部分别作为两个参数处理。在纹理计算时,分别对0~0、45~0、90~0和135~0四个方向的能量、惯性矩、熵、相关性和局部平稳五个参数计算均值和标准差作为特征参数。3、针对高维数据会造成算法空间复杂度和时间复杂度指数增加的问题,论文采用PCA线性参数降维方法对特征参数计算环节得到的24维参数进行降维,得到一个新的4维特征参数,有效解决了高维参数对识别速度和存储容量的影响。4、论文利用神经网络的RBF和BP网络结构对强化木地板表面质量检测进行了分类器设计,并针对特征提取阶段得到的24维和4维特征参数进行了分类识别实验。通过比较得出RBF神经网络在强化木地板表面质量检测方面更具有优势。更多还原

【Abstract】 In accord with the national sustainable developing policies,laminated flooring is growing rapidly and having enormous room for future growth.The surface quality is an integrated-index which directly influences the application of laminated flooring.However,in domestic laminated flooring production line,outward appearance quality inspection still depends on human visual.Thus,this paper brings in the research of method for surface quality detection of laminated flooring which basing on machine vision.The studying subject of this paper is a certain laminated flooring produced by former Beijing Kenuo Senhua floor factory,and mainly research on its most common four defects which are frosting, bare substrate,dirt and tearing of impregnated paper.The method for surface quality detection of laminated flooring basing on machine vision includes the image fetching and segmentation,Feature parameter calculating and extraction,and classifier designing.The mission of segmenting is that the target defects from the complicated wood-grain background should be achieved during the stage of image segmentation.And then calculating based on analyzing the laminated flooring defects to find out the feature parameter which could reflect the qualified and the defected flooring’s attribute.At this stage, the parameter need to be dimensions reduced as well,in order to make sure the parameter is distinctive, reliable,independent and modicum,and reduce the complexity of classifier meanwhile.When coming to the classifier designing,we want to create one kind of classifier which can assign different identifiers to the qualified floor and defective floor.This paper did a systemic study for Surface Defects Detection of laminated flooring.The main contents are as follow:1、At the stage of image segmentation,this paper puts forward three image segmentation methods: Image segmentation based on maximum between-cluster variance;Image Segmentation by Ant Colony Algorithm in two-dimension space;Image Segmentation based on Genetic Algorithm and the maximum entropy,and compares their explicabilities on detecting laminated flooring with each other.The segmentation based on OTSU is not applicable for light color defects including frosting;on contrary,the segmentation based on Ant Colony Algorithm is not applicable for dark color defects including bare substrate,dirt and tearing of impregnated paper defects;The method based on Genetic Algorithm has good segmentations for dark and light color defect,so,the method is applicable for surface image segmentation of Laminated flooring.2、In the research of feature extraction,this paper advances that fetching out the color and grain features of images can express the character of floor based on analyzing the sampled laminated flooring surface image.In the calculation of the color parameter,the HIS three-dimensional space is reduced to one-dimensional space by weighting and summing.Then the first,second and third moment of color moment are calculated in the one-dimensional space.The real and imaginary parts of the third moment are conducted as two features.In the calculation of the texture parameter,the five fetures of energy, moment of inertia,entropy,correlation and local stationarity at four angle directions of 0~0,45~0,90~0 and 135~0 should be calculated and averaged,the standard deviation is resulted as well.3、Targeting the problem that the high-dimensional data would increase the complexity degree of algorithm space and time,this paper use PCA linear parameter to reduce the twenty-four-dimensional parameter result,and get a new four-dimensional feature parameter,which effectively solve the recognizing-speed and storage-capacity problems caused by high-dimensional data.4、Paper use RBF and BP neural network structure to progress the classifier design.Then feature parameters of the twenty-four-dimension and the four-dimension are classifier respectively by the two networks.The result shows that the RBF network has more advantage than the BP network in detection of surface defects for laminated flooring after comparing these two networks.更多还原