【作者】 尹建新；

【导师】 计时鸣；

【作者基本信息】 浙江工业大学 ， 机械电子工程， 2007， 硕士

【摘要】 板材是木材应用需求量最大的品种，板材表面质量是评定板材质量的重要指标之一。随着木材加工业向机械化、自动化的大规模生产方向发展，人们对板材的加工质量，尤其是表面缺陷给予了越来越多的重视，因而表面缺陷检测技术变得越来越重要。本文基于机器视觉理论对木材表面缺陷进行了深入研究，结合数字图像处理技术和人工神经网络模式识别技术，研究了木材表面缺陷图像预处理、特征提取、模式识别问题，以Visual C＋＋程序设计环境，开发了用于检测板材表面缺陷的定位和识别等图像处理算法。图像预处理是检测的第一步，对图像缺陷特征的正确提取是非常关键的。论文通过三种方式对图像的灰度直方图进行分析统计：(1)对每一像素作256级灰度直方图分析；(2)对4×4像素块作256级灰度直方图分析；(3)4×4像素块作16级灰度直方图分析。特征提取直接影响木材缺陷检测系统的识别率。论文首先从灰度直方图中根据是否有颜色突变来判断图片是否存在缺陷，缺陷图片在直方图中表现出双峰特征，通常次波峰即为缺陷部位，但这不是绝对的。若直方图曲线只有一个波峰，则可能是正常木材图片。经过实验统计，当次波峰值比主波峰的值大于1／10时，次波峰即是代表缺陷颜色。实现了缺陷检测的第一步，即把图像分为有缺陷和无缺陷两类。基于人工神经网络的模式识别具有对数据类型和分布函数没有限制、容忍度更高等优点，相适应于木材表面缺陷的复杂性，有很好的应用前景。论文以缺陷灰度均值、缺陷灰度方差和缺陷形状作为缺陷类型识别的特征量为输入，缺陷类型为输出，构建了系统的BP网络系统模型。论文以4种缺陷类型为输出，选用LMS对BP神经网络进行训练，对设计的神经网络系统进行了检测，实验结果表明系统的平均识别率为97％，证实了所设计系统的可行性和有效性。更多还原

【Abstract】 Saw-timber is most needed in wood application, and the surface of saw-timber is one of the important factors to assess wood quality. With large-scaled mechanization and automation of wood processing, people begin to attach more and more importance to manufacturing quality, especially surface defect. Therefore, surface defect detection is becoming more and more important.Based on the theory of computer vision, a research on defect detection of the wood surface is made in paper. Image preprocess, feature extraction and pattem recognition of wood surface defect are also studied by means of digital image processing techniques and neural networks principles of pattern recognition technology. An arithmetic of image disposal was compiled for disfigurement detection to orientate and recognize in the environment of VC++.Image preprocess is the first step for detection, which is vital to the correct extraction of the defection feature. The grey-historam is analyzed through three methods in paper: (1) 256 grade grey-historam statistics for every pixel. (2) 256 grades grey-historam statistics for every 4×4 block pixels. (3) 16 grades grey-historarn statistics for every 4×4 block pixels.Extraction of characteristic quantity will directly affect the distinguishing ratio of wood disfigurement detection system. The paper firstly judged whether there existed defects in the pictures according to the color break in grey-historam. Defect pictures have double peaks in grey-historarn. Generally speaking, the subordinate peak is just the defect part. But it is not absolute. If there is only one peak in the grey-historan, it means that the picture is normal. Experiments show when the value of subordinate peak is 1/10 greater than host peak value, subordinate peak indicates defect color. It would realize the first step for wood detection. and the images are divided into two types: defect images and normal images.Because there is superior tolerance and no limit to data type and distributing function in pattern recognition of neural network, it has a good prospect in application so as to adapt to the complexity in wood surface defect. Using three character values including gray average value, gray variance and defect shape as input value, ten defect types as output. This paper posed BP neural network system model. The designed network system is tested by chosing four type of defects, and LMS arithmetic to train BP neural network. The result shows the average ratio of recognition is 97 percent. The system is viable and available.更多还原