CASIA OpenIR  > 毕业生  > 博士学位论文
精密零件视觉外观检测方法研究
孙佳1,2
学位类型工学博士
导师乔红
2018-05-28
学位授予单位中国科学院研究生院
学位授予地点北京
关键词摄像机标定 表面瑕疵检测 自动检测系统 卷积神经网络 零件测量
摘要精密零件在众多工业领域都有非常重要的应用,如消费电子、医疗器械、航天航空、汽车船舶等等。这些精密零件的外观质量直接影响着最终产品的功能及可靠性。随着我国智能制造战略的推进,对精密零件生产的质量要求也越来越高。在很多应用领域都要求对精密零件外观进行100%全检,因此,对大批量生产精密零件的外观检测效率及检测效果均提出了更高的要求。
目前,基于视觉的精密零件外观检测方法在工业领域中应用广泛,研究人员提出了多种针对不同目标的外观瑕疵检测算法及相应的检测系统。但是在实际应用中仍然存在很多亟需解决的问题,如检测准确率低、样本数量有限、算法泛化能力差等。本文针对精密零件外观检测问题中的视觉系统标定、快速零件检测、表面瑕疵检测等关键技术展开研究,主要的工作和贡献有:
(1)针对精密零件外观检测视觉系统中摄像机内参数及手眼关系的标定问题,分别研究了基于标定板的视觉系统标定方法及自标定方法,并提出了基于两特征点的高效机器人视觉自标定方法。所提出的基于两特征点的自标定方法仅选取空间中两个特征点作为标定参考点,并将机器人视觉系统中的摄像机内参数的标定与机器人及摄像机的手眼标定过程结合在一起,通过控制机器人做三次平移运动和两次旋转运动即可线性求解出所有参数。通过实验验证,该方法在图像噪声等级为σ=5像素时,摄像机内参数的相对误差小于0.06%,手眼参数的相对误差小于2%;在机器人运动误差等级为0.1时,摄像机内参数的相对误差小于0.14%,手眼参数相对误差小于3%。所提出的基于两特征点的自标定方法在减少标定累积误差的同时,简化了机器人视觉系统的标定过程,从而提高了视觉系统的标定效率和灵活性,减少了对标定板等标志物的依赖。
(2)针对精密零件外观检测系统中,对大批量生产的精密零件快速检测和捕捉的实时性及准确性要求,将人类的视觉注意机制引入到零件检测中,提出了基于视觉显著性检测的目标检测算法,将自上而下的特征提取与自下而上的显著性检测相结合,实现目标精密零件的快速精准定位。该方法根据离线先验知识对目标零件进行特征提取并创建样本模板,再对实时采集的待检精密零件图像进行自下而上的显著性检测缩小搜索范围,进而快速锁定检测目标。通过参数分析实验及对比实验可见,本文提出的基于视觉显著性的精密零件检测算法运行时间在4ms左右,当合理设置匹配阈值时,检测成功率可以达到100%。本文提出的基于视觉显著性的精密零件检测算法运行时间短,检测成功率高,能够满足实际应用中的效率与准确性要求。
(3)针对精密零件表面瑕疵分割和提取困难等问题,提出了基于图像重构集合和深度卷积网络的精密零件表面瑕疵检测方法。该方法将训练数据集中的精密零件图像进行自适应多尺度及随轮廓局部提取后,创建重构图像集合,再进行卷积神经网络的训练。所提出的基于图像重构的卷积神经网络能够同时实现多种不同类型瑕疵的有效检测。该方法在精密零件的垫伤、划痕和麻点瑕疵的检测中,均达到了97%以上的检测效果。同时,为了解决缺少带标注的精密零件样本库的问题,本文采集了一万幅精密零件的外观样本图像并对其瑕疵类型进行标注,创建了精密零件表面瑕疵样本数据库。为后续精密零件表面瑕疵检测技术的研究提供了重要依据。
(4)针对小样本精密零件的检测问题,提出了基于合成样本学习的精密零件检测方法,设计了基于合成样本的精密零件视觉外观检测系统。基于合成样本学习的精密零件检测方法利用少量真实样本作为种子样本创建出大量合成样本,对卷积神经网络进行精细训练。有效的解决了实际应用中仅有小样本数据进行检测网络训练的问题。通过实验验证,在垫伤、划痕和麻点的种子样本量分别为三十个时,所提出的方法对相应瑕疵类型检出率分别可以达到98.2%、99.1%及100%。结合前文所研究的内容,针对小尺寸精密零件的自动检测问题,设计了基于合成样本的精密零件视觉外观检测系统。该系统能够实现大批量生产的小尺寸精密零件全自动上料、快速捕捉、表面瑕疵检测以及关键尺寸测量等任务。该方案与现有检测手段相比,大大提高了检测效率,能够实时检测多种瑕疵。
最后,总结了本文的研究成果,并对后续的研究工作进行分析和展望。
其他摘要Precision workpieces are widely used in many industrial areas, such as consumer electronic products, medical instruments, aeronautic vehicles, etc. The quality of work-piece surface directly affects the functionality and reliability of the products. With the need of large scale equipment manufacturing and strategic scientific research, quality demands for precision work-piece are improving. Especially for some work-pieces, 100% of products are required to be checked. Meanwhile, the requirement for efficient detection and measurement for such a large quantity of items is increasing constantly.
At present, the method of work-piece detection based on vision is widely used in the field of industry. Many researchers have proposed a variety of algorithms and the corresponding systems for different surface defect detections. However, there are still some problems to be solved in practical application, such as low detection precision, lack of sample capacity, low generalization ability. This dissertation focuses on the key technologies on the surface detection of precision work-piece, such as, calibration of vision system, work-piece recognition, defect detection and defect classification, and so on. The main work and contributions are as follow:
(1)To solve the problem of camera and hand-in-eye system calibration in precision work-piece detection system, a novel and effective self-calibration approach for robot vision is presented, which can effectively estimate both the camera intrinsic parameters and the hand-eye transformation at the same time. The proposed calibration procedure is based on two arbitrary feature points of the environment, and three pure translational motions and two rotational motions of robot end-effector are needed. From the results of experiment, when the distortion of image noise is σ=5 pixels, the reletive errors of intrinsic parameters are less than 0.06%, and the rotational and translational motion parameters of hand-eye relation are less than 2%. When the distortion of motion noise is 0.1, the reletive errors of intrinsic parameters are less than 0.14%, and the rotational and translational motion parameters of hand-eye relation are less than 3%. The proposed algorithm has been verified by simulated data with different noise and disturbance. Because of fewer feature points and robot motions needed, the proposed method greatly improves the efficiency and practicality of the calibration procedure.
(2)To improve the reality and accuracy of locating and capturing precision work-piece in automatic detection system, a fast and accurate work-piece detection algorithm is proposed based on top-down feature extraction and bottom-up saliency estimation. Firstly, a top-down feature extraction method based on the prior knowledge of work-piece is presented, in which the contour of a work-piece is chosen as the major feature and the corresponding template of the edges is created. Secondly, a bottom-up salient region estimation algorithm is proposed, where the image boundaries are labelled as background queries, and the salient region can be detected by computing contrast against image boundary. In addition, strategies such as image pyramids and a stopping criterion are adopted to speed-up the algorithm. From the results of experiment, the runtime of proposed algorithm is about 4ms, and precision work-pieces can be 100% detected successfully when a appropriate threshold is set. Experiments and results demonstrate the effectiveness of the proposed method.
(3)To solve the problem of surface defect segmentation and extraction, a new detection method based on multi-scale and with contour reconstruction for precision work-pieces is proposed. In this method, the training images are reconstructed by automatic multi-scale extraction and with contour local extraction. A new data set based on reconstruction is established. Then, the convolutional neural network trained by the new data set can realize many kind of surface defects detection at the same time. Experiments on the above methods have been carried out. The results show that the success rate of defect detection by proposed algorithm is above 97%. Meanwhile, in order to solve the problem of lack of precision work-piece data set, over ten thousand precision work-pieces images are collected and labeled, which will be a strong support for the following research of precision work-piece surface defect detection.
(4)To solve the problem of lack of sample set, a surface defect detection algorithm based on synthetic sample for precision work-piece is proposed. A lot of synthetic samples are created from a small number of real precision work-pieces, and a convolutional neural network based on the created synthetic samples are proposed. From the results of experiment, when 30 real precision work-pieces is adopted, the detection results by the proposed algorithm can reach 98.2%, 99.1% and 100% for indentation defect, scratch defect and spot defect, respectively. Combining the above proposed technologies, a real-time detection and measurement system based on man-made samples is proposed for small size precision work-pieces. This automatic system embedded with the proposed detection algorithm based on small samples is designed to pick out defective work-pieces without any manual auxiliary. In addition, the calibrate method for vision system with telecentric lens is discussed, and the dimension of the workpieces is measured.
Finally, the results of the research a re summarized, and the future work is analyzed and prospected.
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/20980
专题毕业生_博士学位论文
作者单位1.中国科学院自动化研究所
2.中国科学院大学
推荐引用方式
GB/T 7714
孙佳. 精密零件视觉外观检测方法研究[D]. 北京. 中国科学院研究生院,2018.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
精密零件视觉外观检测方法研究-签字版.p(4689KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[孙佳]的文章
百度学术
百度学术中相似的文章
[孙佳]的文章
必应学术
必应学术中相似的文章
[孙佳]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。