SIA OpenIR  > 智能检测与装备研究室
基于机器视觉的木材在线分级方法研究
Alternative TitleResearch on on-line grading method of wood based on machine vision
孙棪伊1,2
Department智能检测与装备研究室
Thesis Advisor陈帅
Keyword机器视觉 特征提取 分类器 深度学习 分级
Pages75页
Degree Discipline控制工程
Degree Name硕士
2019-05-17
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract本课题以松木板材为研究对象,研究了木材表面的图像预处理、特征提取和自动分级方法,在此基础上建立了木材表面检测软硬件平台,开发出以机器视觉为核心技术的松木板材在线分级系统。主要内容如下:(1)板材经过加工后表面存在毛刺、机器压痕和亮度不均匀等较多的干扰,为提高所采集图像的可读性,减小干扰对后续特征提取的影响,首先研究了图像预处理相关技术,具体包括图像的空域滤波方法、频域滤波方法和图像增强方法。针对毛刺产生的噪声干扰,研究对比了平均邻域法、中值滤波与高斯滤波三种滤波方法,结果表明高斯滤波效果最好,既可以保留纹理的边缘信息,还可以过滤掉毛刺等产生的噪声干扰;针对在机器加工过程中产生的规律性网格状压痕,研究了频域去噪技术,根据压痕在倾斜方向上有明显周期性分布的特点,利用傅里叶变换方法有效地减轻了压痕对特征提取和分级造成的干扰;针对亮度不均匀的图像表面,研究对比了直方图均衡化方法和自适应对比度增强方法,结果表明自适应对比度增强方法效果最理想,突出了木材纹理与背景的区别。(2)针对松木板材的纹理特征提取,研究了灰度共生矩阵、马尔科夫随机场和局部二值模式纹理提取方法。针对灰度共生矩阵方法,选取了二阶矩(能量)、同质性、相关性、对比度、边缘宽度、偏差、熵、灰度中心这八个特征。马尔科夫场可以很恰当的表现纹理结构单元间的相互作用的大小和方向,对于松木板材的纹理识别,像素之间有序的空间性适合用马尔科夫场来描述。利用已知板材条纹的线性特征,提出了一种针对条纹纹理特征提取的线性LBP算子。针对颜色特征提取,研究了颜色直方图和基于混合高斯模型的蓝变区域面积提取方法,用于提取颜色特征和检测蓝变缺陷。由于条纹板材具有周期性分布和局部结构自相似度高的特点,研究了板材表面分区域相似度特征提取和度量方法。(3)针对分级方法,研究了支持向量机、多层感知机和深度学习方法,实验结果表明,分类器对木板的分类准确率达到96%。对比本文提到的三种分类方法,考虑到对于木板图像的适用性,再根据实验结果,样本量不够且相似程度高的木板图片在进行深度学习训练并分类的时候遇到的困难较多,时间成本也照传统分类器相比投入要大。针对于木板图像而言,传统的分类器更加准确和高效。(4)最后,根据课题的具体检测需求,设计了松木板材自动分级系统硬件检测平台,完成了在线视觉检测系统软件设计,验证了分级算法的可行性,实现了松木板材的在线分级。
Other AbstractThis paper takes the pine board as the research object, studies the image preprocessing, feature extraction and automatic grading methods of the wood surface, establishes the software and hardware platform of the wood surface detection on this basis, and develops the pine board online grading system with machine vision as the core technology. The main contents are as follows: (1) The board surface after machining burrs, indentation machine and brightness non-uniformity more interference, in order to improve the readability of the images collected, reduce the influence of interference for the follow-up feature extraction, image preprocessing technology is studied, and the first concrete including image filtering method, frequency domain noise method and contrast enhancement method. Aiming at the noise interference caused by burr, the average neighborhood method, median filter and gaussian filter are compared. The results show that the gaussian filter has the best effect, which can not only retain the edge information of texture, but also filter out the noise interference caused by burr. Aiming at the regular grid-shaped indentation produced in the machining process, the frequency domain denoising technology is studied. According to the characteristic that the indentation has obvious periodic distribution in the inclined direction, the interference caused by the indentation to feature extraction and classification is effectively reduced by using the fourier transform method. For the image surface with uneven brightness, histogram equalization, adaptive histogram equalization and adaptive contrast enhancement methods are studied and compared. The results show that the adaptive contrast enhancement method has the best effect and highlights the difference between wood texture and background. (2) According to the texture feature extraction of pine board, the extraction methods of grayscale symbiosis matrix, markov random field and local binary mode texture were studied. For the above gray symbiosis matrix method, eight characteristics are selected: second moment (energy), homogeneity, correlation, contrast, edge width, deviation, entropy and gray center. Markov field can properly represent the size and direction of the interaction between texture structural units. For the texture recognition of pine board, the ordered space between pixels is suitable to be described by markov field. A linear lbp operator is proposed for the extraction of fringe texture features. The color histogram and the main color extraction method were studied to extract the main color and detect the blue defect. Because of the characteristics of periodic distribution and high self-similarity of local structure, the feature extraction and measurement method of similarity degree are studied. (3) Two classifier methods, support vector machine (svm) and multi-layer perceptron (mpr), and deep learning method, are studied for classification methods. The experimental results show that the classifier's classification accuracy of wood boards is up to 96%. Compared with the three classification methods mentioned in this paper, considering the applicability of the board image, it is concluded from the experimental results that the board image with a small sample size and a high degree of similarity has many difficulties in carrying out deep learning training and classification, and the time cost is higher than that of traditional classifier. Therefore, the traditional classifier is more accurate and efficient for wood image. (4) The software and hardware testing platform of the automatic grading system for pine wood panel was designed to realize the online grading of the panel. Experimental results show that the wood classification method based on machine vision can effectively distinguish the wood texture. Both classification accuracy and defect identification rate can reach more than 93%. This subject provides an important theoretical basis for automatic online detection and grading of sheet materials.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/25201
Collection智能检测与装备研究室
Affiliation1.中国科学院沈阳自动化研究所
2.中国科学院大学
Recommended Citation
GB/T 7714
孙棪伊. 基于机器视觉的木材在线分级方法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2019.
Files in This Item:
File Name/Size DocType Version Access License
基于机器视觉的木材在线分级方法研究.pd(5253KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[孙棪伊]'s Articles
Baidu academic
Similar articles in Baidu academic
[孙棪伊]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[孙棪伊]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.