金属工件表面瑕疵检测技术的研究与开发

发布时间：2018-07-12 17:57

本文选题：金属表面 + 瑕疵检测　；参考：《江南大学》2013年硕士论文

【摘要】：金属工件广泛应用，是各种器械不可缺少的部件，随着生产力的发展，用户对其质量也有更高的要求，而表面质量是最直观的体现，要求也往往更加严格。金属工件表面瑕疵使工件外表不美观的同时，更恶劣的是会影响工件的使用性能，使产品的安全性降低，必须在出厂前剔除。目前国内大部分厂家还是采用人工目测法检测，抽检率低，检测速度慢，检测结果易受检测人员主观因素影响，缺乏一致的、科学的指导，各金属制品企业亟需先进的表面瑕疵检测技术和设备。国外设备和技术，不仅价格昂贵，维护费用高，而且还没有自主知识产权，这些都迫使我们要开发出符合企业需要的自动化瑕疵检测设备和技术。本文以实验室现有各种金属工件为检测研究对象，结合机器视觉、图像处理和模式识别等技术完成对本课题的研究。（1）分析了表面瑕疵检测流程中预处理、图像分割、特征提取阶段现有的一些算法，对其原理和算法实现进行研究，并实验论证，为下一步的工作提供充分的理论基础和技术支持。（2）以轴承为研究对象，提出一种基于机器视觉技术的轴承防尘盖表面瑕疵检测方法，从硬件环境的搭建，到软件算法的实现进行了详细说明。采用蓝色同轴光源作为检测系统所用光源，克服金属反光；采用最小二乘法拟合轴承外圆，根据轴承型号比例分割出防尘盖区域，然后利用Otsu阈值分割和Roberts边缘提取处理图像，再与模板轴承比较，求出相差角度，由此将防尘盖字符、非字符区域分离，两部分是否存在瑕疵分开判别，互不干扰。（3）以铁氧化物--磁瓦为研究对象，分析其表面瑕疵的类型和特点，运用纹理分析的方法实现特征的提取。通过对Gabor滤波器参数表达式的研究，构造了不同尺度、不同方向的Gabor滤波器组，并针对磁瓦表面瑕疵特点对Gabor滤波器组进行了改进，为了去除数据相关性和冗余性，运用主成分分析法和独立成分分析对提取到的特征进行了降维。（4）对BP神经网络和支持向量机的基本原理和算法实现方法进行了研究，针对BP神经网络存在的不足，利用附加动量和变学习率学习的方法进行改进；针对支持向量机核参数c和惩罚因子g选取困难，采用网格法和K-CV法对其实现寻优。最后用磁瓦表面瑕疵数据对两种分类器的分类效果进行实验比较和结果分析。通过实验证明，本文提出的轴承防尘盖表面瑕疵检测方法，，检测系统采集到的轴承图像清晰，瑕疵检测算法正确率在96%以上，可实时的完成轴承防尘盖表面瑕疵自动检测。通过改进的Gabor滤波器组，实现了磁瓦表面瑕疵的特征提取，采用PCA，ICA分析法实现了特征降维，采用本文算法对磁瓦表面瑕疵进行分类，总体正确率可以达到93%以上，为表面瑕疵检测分类提供了一种新方法。
[Abstract]:Metal workpieces are widely used and are indispensable parts of various instruments. With the development of productivity, users have higher requirements for their quality, and the surface quality is the most intuitive embodiment, and the requirements are often more stringent. The surface defects of metal workpiece make the appearance of workpiece unattractive, and at the same time, it will affect the performance of workpiece and reduce the safety of product, so it must be eliminated before leaving the factory. At present, most domestic manufacturers still use manual visual testing. The sampling rate is low, the detection speed is slow, the test results are easily influenced by the subjective factors of the examiners, and they lack consistent and scientific guidance. All metal products enterprises need advanced surface flaw detection technology and equipment. Foreign equipment and technology, not only expensive, high maintenance costs, but also do not have independent intellectual property rights, which force us to develop automatic defect detection equipment and technology that meet the needs of enterprises. In this paper, we take all kinds of metal parts in the laboratory as the detection object, and finish the research on this subject with machine vision, image processing and pattern recognition technology. (1) analyze the pretreatment and image segmentation in the process of surface defect detection. At the stage of feature extraction, some existing algorithms are studied, and the experimental results are presented to provide sufficient theoretical basis and technical support for the next work. (2) taking bearing as the research object, This paper presents a method for detecting the surface defects of bearing dust-proof cover based on machine vision technology, from the construction of hardware environment to the realization of software algorithm. The blue coaxial light source is used as the light source of the detection system to overcome the metal reflection, the least square method is used to fit the outer circle of the bearing, and the dust-proof cover area is segmented according to the bearing type ratio, and then the image is extracted and processed by Otsu threshold and Roberts edge. Then compared with the formwork bearing, the angle of difference is calculated, and the dust-proof cover character, the non-character area and the defect of the two parts are separated. (3) the iron-oxide magnetic tile is taken as the research object. The types and characteristics of surface defects are analyzed, and the feature extraction is realized by texture analysis. By studying the expression of Gabor filter parameters, Gabor filter banks with different scales and directions are constructed, and Gabor filter banks are improved to remove data correlation and redundancy. Principal component Analysis (PCA) and Independent component Analysis (ICA) are used to reduce the dimension of the extracted features. (4) the basic principles and algorithms of BP neural network and support vector machine are studied, and the shortcomings of BP neural network are pointed out. The method of learning with additional momentum and variable learning rate is improved, and the kernel parameter c and penalty factor g of support vector machine are difficult to select, and the mesh method and K-CV method are used to optimize them. Finally, the classification effects of the two classifiers are compared and analyzed by using the surface defect data of the magnetic tile. It is proved by experiments that the method proposed in this paper is clear in the image of bearing collected by the detection system, and the correct rate of defect detection algorithm is over 96%, which can be used to detect the surface defects of the dust proof cover in real time. Through the improved Gabor filter bank, the feature extraction of the surface defects of the magnetic tile is realized, and the feature dimension reduction is realized by using the PCACICA analysis method. The algorithm of this paper is used to classify the surface defects of the magnetic tile, and the overall correct rate can reach more than 93%. It provides a new method for the classification of surface defect detection.
【学位授予单位】：江南大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP274

【参考文献】