【作者】 谢永华；

【导师】 王克奇；

【作者基本信息】 东北林业大学 ， 机械设计及理论， 2013， 博士

【摘要】 木材表面缺陷检测技术是计算机视觉与模式识别相交叉的多学科技术,该技术具有较高的应用价值,被广泛应用在木材生产及其深加工等领域.。本文主要以木材的死节、活节和虫眼三种常见缺陷为研究对象,对木材的缺陷图像分割和模式识别方法进行了深入的研究。主要内容包括：木材表面缺陷图像分割、特征提取、缺陷类型识别等问题。图像分割是木材缺陷识别的首要问题。本文对传统边缘检测算法进行了介绍,针对分水岭过度分割不足,结合数学形态学提出了基于形态学梯度的分水岭分割方法,并将其应用在木材缺陷的检测中；针对在计算分形参数而产生的边缘效应的问题,提出了增维矩阵的计算方法；针对木材缺陷的颜色特征,结合模糊聚类算法,提出基于颜色矩的图像分割方法。同时采用具有强大运算功能的数学形态学工具,对分割后图像进行了后处理工作,加强了分割图像的可视性,确保了缺陷图像的分割精度,为缺陷的识别工作奠定了基础。为保证木材缺陷识别结果的可靠性,特征量的选取是模式识别中至关重要的环节。本文对木材缺陷分割图像进行Tamura纹理、灰度共生矩阵、小波多分辨率分形维的特征提取,分别选用了BP神经网络和支持向量机分类器进行分类识别。其中,以多分辨率分形维作为BP神经网络分类器的输入,无论采用何种训练函数,分类的准确率均达到92.67%；在支持向量机分类器中,Tamura纹理与灰度共生矩阵联合的10个参数的识别准确率高达96.67%,多分辨率分形维的识别准确率也高达94.00%,准确率均高于BP神经网络分类器。试验结果表明：(1)支持向量机在对木材缺陷图像进行分类,特别是它在解决小样本、非线性及高维模式识别中表现出许多特有的优势。(2)运用数字图像处理技术,根据木材表面缺陷图像的颜色特征来解决木材表面缺陷的分割,根据缺陷图像的纹理特征来解决识别等问题,是行之有效的途径。更多还原

【Abstract】 The wood surface defect detection is an interdisciplinary technology, which has a higher value in the field of timber production and deep processing. In this paper, burrow,dead knot,slipknot are the most three common wood defects for the study, conducted in-depth research on the wood surface defect pattern recognition method. The main contents include wood surface defects image segmentation, feature extraction and defect type recognition problem and so on.Image segmentation is the most important issue of the wood defect recognition. This article introduces the traditional edge detection algorithm. For watershed over-segmentation, wood defect image segmentation based on morphological watershed segmentation method has been proposed. Improvements have been made for the deficiencies of edge effect based on the fractal parameters wood surface defects segmentation method, it has put forward increased dimensional matrix calculation method. For timber defects color features, clustering segmentation method based on color moments has been proposed combined with fuzzy C-means clustering algorithm. While using mathematical morphology tools with powerful computing capabilities, processing the segmented image to strengthen the visibility and the integrity of the divided image and improve the accuracy of the defect extraction.The extraction of the characteristic quantities directly affect the recognition rate of the timber defect detection system. The feature extraction of wood defects image segmentation using Tamura texture,GLCM,wavelet multi-resolution fractal dimension were selected for classification with BP neural network and support vector machine classifier, Which takes multi-resolution fractal dimension as the input of BP neural network classifier, regardless of the training function, the classification accuracy rate reaches92.67%. In support vector machine classifier, Tamura texture and gray symbiotic matrix combined10parameters recognition accuracy rate is up to96.67%. The recognition accuracy rate of Multi-resolution fractal dimension is up to94%, which is higher than the BP neural network classifier.The test results show that,(1)support vector machines works in wood defect image classification, in particular, it demonstrates many unique advantages in solving small sample, nonlinear and high dimensional pattern recognition.(2) Using digital image processing technology, they are identification and effective ways to resolve issues to take segmentation based on the color characteristics of the wood surface defect images to solve the segmentation of the wood surface defects and based on defect image texture features.更多还原