基于小波理论的高光曲面弱特征缺陷辨识及虚拟现实展示

高精度轧辊是生产军用导线包覆箔、燃料电池极板等高精度产品必不可少的部件。然而目前国内产品相较于国外，轧辊的关键技术指标上尚不能满足要求。与此同时，高光泽轧辊表面会出现一些区别于以往的弱特征缺陷，这些缺陷尚没有明确的定义且无法经过传统的表面检测方法检出，但却极易复映到被加工件表面，这会使产品质量不达标准。在行业研究中，机器视觉检测技术被认为是最为有效监控表面质量缺陷的方法。本文将按照机器视觉表面缺陷检测的方法为主体思路，并结合应用前景极为广泛的虚拟现实技术，展开如下研究内容： 首先，以获取高质量图像为切入点，搭建机器视觉拍摄平台，对系统中关键组成部分给出选型依据及品牌型号。拍摄出轧辊表面常见弱特征缺陷，对常见缺陷进行定量化分析与描述。 其次，对前述缺陷图像特点加以分析，结合不同缺陷特征提取问题研究成果，论证小波方法在纹理特征提取中具备的独特优势。讨论 Gabor 滤波器的设计步骤和构建原则，提取出可以应用于后续缺陷判别的特征向量。 再次，不同种类的特征中，有可能存在相关度较高成分。同种缺陷特征中也存在着相关性，造成信息冗余的同时增加分类时间、降低分类效率。本文在分析核主成分分析的方法，对向量中的有效信息进行“提取”，实现了特征向量的相互独立与降维；此后根据本课题的外在需求提出应用支持向量机的多类分类方法，在有限样本情况下科学设计交叉验证方式进行实验，实践证明此方法兼顾了准确率与效率，具备在后续的应用中得以推广的价值。 最后，鉴于存在缺陷特征不明显，并不能被识别和有效分类的图像这类情况，轧辊缺陷出现的位置和缺陷位置的图像尚未进行融合，生产管理方需要一种直观有效的监控方法掌握生产过程中轧辊一手信息等客观需求，在综合已有研究成果的有益经验基础上，展开了对虚拟现实方法的分析与研究，利用终端操作引擎 Hololens 和 Visual studio 实现对缺陷的虚拟现实三维呈现，在现实体验过程中发现了若干仍需优化的技术细节并进行了改进。 本文着眼于目前行业内尚无明确定义的高精度高光泽轧辊表面弱缺陷特征，结合机器视觉成熟研究成果，对其中各关键环节进行了分析与设计。在样本有限的条件下，通过合理设计实验方案，验证了该系统应用于后续工厂环境中的可能性。此外本文中使用的三维虚拟现实方法也将为将来的大规模推广做出了有益尝试。

High precision roller is an indispensable part for producing high precision products such as coated foil and fuel cell plate. However, compared with foreign countries, the key technical indexes of roll can not meet the requirements at present. High-glossy roll surface, meanwhile, existing weak character defects different from the previous, which has no clear definition and couldn’t be detected by the traditional method, is easy to rerun to the work piece surface, causing the deterioration of the quality of the product. Machine vision detection technology is regarded as the most effective method to monitor surface quality defects in industry research.This paper will carry out the following research contents according to the virtual reality technology with a wide range of application prospects. First, to obtain high-quality images, we set up the machine vision shooting platform, then give the selection basis and brand model for the key components in the system. The defects of common weak features on the surface of roll were taken and quantitative analysis and description of common defects were carried out. Second, the characteristics of the aforementioned defect images are analyzed, and the research results of different defect features are combined to demonstrate the unique advantages of wavelet method in the extraction of texture features.The design steps and construction principles of Gabor filter are discussed and the eigenvectors can be applied to the subsequent defect discrimination. Third, there may be a higher degree of relevancy among different kinds of features.In the same defect, the correlation is also found, which causes the information redundancy to increase the classification time and reduce the classification efficiency. In this paper, we analyze the method of principal component analysis to "extract" valid information in vector, realizing the independence and dimensionality of feature vector. Extreme learning machine classification method(ELM) is put forward according to this topic of external demand. Scientific way cross validation experiment methods are designed under the condition of limited samples, then the follow-up practice prove this method not only accurant but also efficient, has the value that can be promoted in the following application. Finally, there are still existing following problems such as characteristics of defect is not obvious, thus can’t be identification and classification of image effectively, roller defects and defect location has yet to image fusion, the production management requires an intuitive and effective monitoring method to master the process of roller production information. According to the above objective requirments, the analysis and research of virtual reality method is carried out, the virtual reality 3D rendering of defects is realized by using the terminal operating engine Hololens and Visual studio.