基于灰度共生矩阵和分层聚类的刨花板表面图像缺陷提取方法  

Research on Defect Extraction of Particleboard Surface Images Based on Gray Level Co-Occurrence Matrix and Hierarchical Clustering

作　　者：郭慧 王霄 刘传泽[2] 周玉成[1,2] Guo Hui;Wang Xiao;Liu Chuanze;Zhou Yucheng(Research Institute of Wood Industry,CAF Beijing 100091;School College of Information and Electrical Engineering,Shandong Jianzhu University Jinan 250101)

机构地区：[1]中国林业科学研究院木材工业研究所,北京100091 [2]山东建筑大学信息与电气工程学院,济南250101

出　　处：《林业科学》2018年第11期111-120,共10页Scientia Silvae Sinicae

基　　金：中国林业科学研究院中央级公益性科研院所基本科研业务费专项资金(CAFYBB2018MB002);泰山学者优势特色学科人才团队(2015162)

摘　　要：【目的】提出一种基于灰度共生矩阵和分层聚类的刨花板表面图像缺陷提取方法,根据缺陷部分与正常部分纹理特征不同,利用分层聚类算法将缺陷分离出来,以解决板面缺陷检测系统中刨花板表面纹理导致缺陷提取不准确的问题。【方法】将刨花板表面灰度图像划分成若干个窗口,使用灰度共生矩阵的统计特征参数对各窗口纹理进行表征,通过分层聚类算法将纹理特征不同区域区分开。首先确定灰度共生矩阵构造因子的取值,包括窗口大小、灰度级、方向和步长,构建出各个窗口的灰度共生矩阵;使用Fisher准则和线性相关性对灰度共生矩阵14个统计特征参数的表征能力进行度量,选取出分类能力强且相关性低的特征构成特征向量,所有窗口的特征向量构成样本集。然后运用BIRCH分层聚类算法对样本集进行聚类,为使聚类结果更准确,同时加快计算速度,提出一种优化策略,绘制样本集均值和统计直方图,将其波峰数量作为理想的类别数量,当聚类产生的类别数量大于理想类别数量时,将聚类结果中距离近的簇合并,解决聚类精度过高而导致的过分割问题。最后根据聚类结果,对原图像中各窗口进行标记,提取出缺陷区域。【结果】选择大小为512像素×512像素,带有杂物、油污、胶斑、大刨花和松软5种类型缺陷的刨花板表面图像,使用本研究方法能够准确将缺陷区域提取出来,精确度达92.2%,召回率达91.8%。【结论】基于灰度共生矩阵和分层聚类的刨花板表面图像缺陷提取方法,可解决因刨花板表面纹理导致缺陷提取不准确的问题,为机器视觉板面缺陷检测系统的缺陷度量和识别提供良好支撑。【Objective】 A method for extracting defect regions on image of particleboards surfaces using gray level co-occurrence matrix and hierarchical clustering was proposed in this paper,which separated the targets from images according to the texture differences between defects and normal parts,in order to solve the problem of inaccurate defect segmentation caused by particle board surface texture in board surface defect detection system.【Method】 The surface image was divided into several small windows and the texture of each window was characterized by the statistical textural feature parameters of the gray level co-occurrence matrix.Then, a hierarchical clustering method was used to distinguish the defect windows and the normal windows using the texture feature parameters. Firstly,the values of the four structural factors were chosen to build the gray level co-occurrence matrixes for each window,including the window size,gray level,direction and step. Secondly, the classification ability and correlation of the 14 statistical parametersof textural features for gray level co-occurrence matrix were evaluated using fisher criteria and linear correlation. Bydoing this,the features which have better classification ability and low correlation were chosen. The feature vector of each window was composed of the selected features values and all of the feature vectors constituted a sample set. The sample set was clustered by the BIRCH hierarchical clustering algorithm. In order to obtain an accurate clustering result ,an optimization strategy was proposed in this paper. The histogram of sum average was drawn and the number of wave peaks was counted as the target category quantity. When the number of classes generated by the initial clustering was larger than the target category quantity,the clusters having closer distance were merged, which can avoid the over segmentation caused by high clustering precision. Finally,according to the clustering result ,the windows in the original image were all marked and the defect areas

关 键 词：灰度共生矩阵 纹理特征 分层聚类 BIRCH算法 缺陷检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] S784[自动化与计算机技术—计算机科学与技术]