CASIA OpenIR  > 毕业生  > 博士学位论文
深孔类零部件内部缺陷的视觉检测方法研究
宫新一
Subtype博士
Thesis Advisor徐德
2019-05-29
Degree Grantor中国科学院大学,中国科学院自动化研究所
Place of Conferral中国科学院自动化研究所
Degree Name工学博士
Degree Discipline控制理论与控制工程
Keyword深孔零件 缺陷检测 视觉检测 深度卷积神经网络 目标检测 轮廓提取 形态分析
Abstract

      随着航天技术、光电技术、新能源技术与生物医学工程技术的不断发展,微机电系统、微型靶件、微光学器件等以形状尺寸微小为特征的微机械得到了越来越广泛的应用,而深孔类零件作为其中的基础原件更是尤为重要。因此,对深孔类零件的高精度质量检测提出了越来越高的要求。在实际生产中,深孔零件孔径小、孔深大、内部结构复杂的特点给其质量检测带来了很大的困难。传统的深孔零件检测依赖于人工操作,操作人员虽然能够完成检测要求,但是在检测过程中容易受到人员主观意识的影响,导致检测指标的一致性往往无法保证。同时,大批量生产需要更多的检测人员,也导致了生产成本居高不下。如何实现高性能、高效的深孔类零件质量自动检测与缺陷识别成为了一个重要的研究课题。本文针对深孔类零件质量检测中的关键问题展开研究。论文主要的工作和贡献如下:

1)搭建了一套深孔类零件质量自动检测与分拣平台。深孔零件质量大多采用基于零件物理特性的非视觉方法进行检测,而这些方法无法保证检测的准确性与鲁棒性。相比之下,基于视觉的缺陷检测方法能够有效对零件质量进行检测,然而深孔类零件具有孔径小、孔深大的特点,其内部清晰图像的获取一直以来是个难题。针对该问题,本文首先设计了一套基于工业内窥镜的视觉成像系统,并在此基础上开发了一套完整的零件质量自动检测与分拣平台。该系统克服了常规视觉成像系统难以采集零件内部清晰图像的困难,可以稳定、清晰地获取深孔零件内部图像,为实现深孔类零件质量自动检测奠定了基础。

2)提出了基于先验知识的多目标形态信息提取算法。深孔类零件质量与其内部结构形态直接相关,然而零件内部结构复杂,其图像中存在着大量的阴影与噪声,且图像对比度较低,常规的特征提取算法难以准确获取内部结构形态信息。为了解决该问题,本文首先提取出了一种基于卡尔曼滤波的骨架提取算法。该算法利用先验知识提取零件中关键点,随后通过极值点搜索方法提取图像中属于零件结构的像素点,并利用卡尔曼滤波算法对其位置进行修正,以削弱噪声与阴影对极值点搜索的影响。随后,本文提出了一种基于布朗运动模型的轮廓提取算法,以零件中关键结构轮廓的曲率分布作为先验知识,建立布朗运动模型,计算零件边缘图像中各边缘属于目标轮廓的概率,通过阈值筛选出目标轮廓的候选边缘,最终通过边缘聚类的方式提取断裂的目标轮廓。本文提出的布朗运动模型能够去除图像中的冗余边缘,有效地提升轮廓提取的实时性与准确性。

3)提出了一种无先验知识的多目标形态信息提取算法。深孔零件具有多种尺寸和结构,利用先验知识虽然可以准确提取某种类型零件内部结构形态信息,但是对于不同种类的零件泛化性较差。为了解决不同类型零件中目标形态的提取,提出了一种基于Faster R-CNN的最小外接矩形提取方法。在Faster R-CNN的基础上进行改进,使其检测结果输出从常规的目标包络框变为目标最小外接矩形。随后,将之与基于卡尔曼滤波的骨架提取算法和基于布朗运动的轮廓提取算法相结合,基于所提取的目标最小外接矩形的特征,在无先验知识情况下的对多目标骨架与轮廓进行提取,从而实现了不同类型零件中关键结构形态信息的提取。

4)设计了一种深孔类零件质量检测方法。首先,利用基于Faster R-CNN的最小外接矩形提取算法、基于卡尔曼滤波的骨架提取算法和基于布朗运动的轮廓提取算法对零件中不同的形态特征进行提取。随后,对不同的形态特征进行分析,确定可以直接反映零件缺陷特性的关键因素,最终通过制定相关规则实现零件质量检测与缺陷分类。本文方法在工业现场采集的图像上进行了测试,结果表明,所提出的方法能够有效地对零件质量及缺陷种类进行判定,与现有方法相比,本文所提方法在准确性上具有明显的提升,能够满足深孔类零件质量检测需求。

      最后,对本文中所提出的研究成果进行了总结,并对未来的研究工作进行了分析与展望。

Other Abstract

      With the continuous development of aerospace, optoelectronic, new energy and biomedical engineering, the micro-mechanical systems such as micro-electro-mechanical system (MEMS), micro-targets and micro-optics have been highly developed and found an increasingly wide application. As their basic element, the components with deep hole are especially important. Therefore, the quality inspection of such component is highly demanded. In practical production, due to small diameter, large depth and complex internal structure of the component, its quality inspection is always a difficult problem. The conventional inspection relies on manual operation. Although the operators can implement the inspection task, it is easily affected by their subjective consciousness during the inspection process, and the consistency cannot be guaranteed. At the same time, mass production requires more inspectors, which also leads to high production costs. It has become an important research to achieve automatic quality inspection and defect identification of the component with high precision and efficiency. This paper focuses on the key issues in the quality inspection of the component with deep hole. The main work and contributions are as follows:

(1) An automatic pick-and-place system for the component with deep hole is constructed. Quality inspection of the component is always a technical difficulty. In most cases, the quality is inspected by non-visual methods based on the physical characteristics of the components, and these methods cannot guarantee the accuracy and robustness of the detection. In contrast, the vision-based defect inspection method can effectively inspect the quality of components. However, due to the characteristics of small aperture and deep hole, the acquisition of clear images inside the component has always been a problem. In response to this problem, a visual imaging system based on industrial endoscope is designed, and a complete automatic quality inspection pick-and-place system is developed. The system can overcome the difficulty that the conventional visual imaging system cannot collect clear images inside the component and achieve the acquisition of the component’s internal image stably and clearly. The construction of the system lays a foundation for realizing the automatic quality inspection of components with deep hole.

(2) A multi-objective morphological feature extraction algorithm based on prior knowledge is proposed. The quality of the component is directly related to its internal structure. However, it is difficult for the conventional feature extraction algorithm to accurately extract the internal structure shape information due to shadows, noises and low contrast of the image and the complex internal structure. In order to solve this problem, a skeleton detection algorithm based on Kalman filter is proposed. The algorithm uses prior knowledge to detect key points in the component. Then the pixels belonging to the component structure in the image are extracted through the extreme point searching method. Finally, the point’s position is amended by Kalman filter algorithm to weaken the impact of noise and shadow. Subsequently, a contour extraction algorithm based on Brownian motion model is proposed. The curvature distribution of key structural contours in the component is utilized as a priori knowledge to establish a Brownian motion model. The probability that each edge belongs to the object’s contour is calculated in the edge image of the component. The candidate edges of the contour are filtered out, and finally the fracture contour of the object is extracted by edge grouping method. The Brownian motion model proposed in this paper can remove the redundant edges arising from the noises and shadows in the image and effectively improve the real-time performance and accuracy of contour extraction.

(3) A multi-objective morphological information extraction algorithm without prior knowledge is proposed. Deep hole components have a variety of sizes and internal structures. Although the prior knowledge can be utilized to accurately extract the internal structural shape information of a certain type of parts, the generalization for different types of component is poor. In order to achieve the extraction of shape information in different types of components, a minimum bounding rectangle extraction method based on Faster R-CNN is proposed. Based on Faster R-CNN, the detection result is changed from the regular bounding box into the minimum bounding rectangle. Subsequently, combined with the Kalman filter-based skeleton detection algorithm and the Brownian motion-based contour extraction algorithm, utilizing the extracted minimum bounding rectangle, the objects’ skeleton and contour are extracted without prior knowledge. Sequentially, the shape information extraction of key structural in different types of components is realized.

(4) A quality inspection method of deep hole component is designed. Firstly, the minimum bounding rectangle extraction algorithm based on Faster R-CNN, the skeleton extraction algorithm based on Kalman filter and the contour extraction algorithm based on Brownian motion are adopted to extract different morphological features of the internal structure in the component. Subsequently, different morphological characteristics are analyzed to determine the key factors that can directly reflect the defect characteristics of the component. Finally, the quality inspection of component and defect classification are realized by formulating relevant regulation. The method is validated on the images acquitted in the practical industrial field. The experimental results show that the proposed method can effectively detect the quality of component and the types of defects. Compared with the existing methods, the proposed method achieves significant improvement in accuracy, which can meet the requirements of deep hole components’ quality inspection.

Pages160
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/23925
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
宫新一. 深孔类零部件内部缺陷的视觉检测方法研究[D]. 中国科学院自动化研究所. 中国科学院大学,中国科学院自动化研究所,2019.
Files in This Item:
File Name/Size DocType Version Access License
毕业大论文.pdf(77289KB)学位论文 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[宫新一]'s Articles
Baidu academic
Similar articles in Baidu academic
[宫新一]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[宫新一]'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.