【作者】 刘悦；

【导师】 崔俊宁；

【作者基本信息】 哈尔滨工业大学 ， 仪器科学与技术， 2024， 硕士

【摘要】 基于X射线图像的精密构件内部缺陷检测方法,是航空航天、核工业等关键领域中精密构件的质量检测和评估的关键技术手段。X射线图像包括二维DR(数字成像,Digital Radiography)图像和三维CT(断层扫描,Computed Tomography)图像两种,二维DR图像用来对缺陷进行定性检测,确定缺陷是否存在并粗略估计缺陷的大小,三维CT图像用来对缺陷进行定量测量,确定缺陷的位置和体积等三维特征。国内对基于X射线图像的缺陷检测算法的研究起步较晚,在检测精度和效率上与国外先进水平有较大差距。目前,基于X射线图像的精密构件内部缺陷检测方法主要有以下问题亟待解决:(1)基于二维DR图像的人工特征提取类缺陷检测方法,在低对比度图像场景下难以获得缺陷的高判别性特征,漏检率和误检率较高;(2)多数基于二维DR图像的深度学习缺陷检测网络未考虑缺陷特征的不均衡性,抗干扰能力弱,易受噪声影响,难以实现高检测精度;(3)目前基于单层切片进行的三维CT图像的三维缺陷检测方法,受单层切片检测结果影响极大,容错率低,检测精度有限,且对缺陷三维空间相关性信息的利用率低。针对上述问题,本课题对基于X射线图像的精密构件内部缺陷检测方法进行了研究,主要完成了以下工作:首先,针对人工特征提取类缺陷检测算法难以获得缺陷高判别性特征的问题,设计了一种基于双阈值重构的二维DR图像缺陷检测算法,利用局部动态阈值分割和双阈值重构思想,在局部区域内提取缺陷的高判别性灰度特征,降低缺陷漏检率和误检率。该方法可成功检测出全局阈值分割、灰度梯度提取等人工特征提取类方法难以检测出的缺陷,缺陷漏检率为16.9%,误检率为10.8%,符合预期目标,方便快捷,检测效率高,对数据集依赖性小,适用于对检测效率要求较高且数据集难以获取的二维缺陷检测场景。其次,针对深度学习缺陷检测网络未考虑缺陷特征的不均衡性的问题,设计了一种基于不均衡卷积特征提取的二维DR图像智能缺陷检测算法,构建不均衡卷积特征提取网络,并引入距离注意力机制优化损失函数,对YOLOv8模型进行改进。通过实验验证,改进后模型的缺陷漏检率为5.5%,误检率为5.1%。与原YOLOv8模型相比,缺陷漏检率和误检率分别降低了4.7%和1.8%,与基于双阈值重构的二维DR图像缺陷检测算法相比,缺陷漏检率和误检率分别降低了11.4%和5.7%,缺陷检测精度大幅提高,适用于高检测精度要求的二维缺陷检测场景。最后,针对基于三维CT图像的单层切片进行的三维缺陷检测方法受单层切片检测结果影响大的问题,设计了一种基于模糊聚类和区域生长的三维缺陷检测算法,基于模糊聚类思想直接对三维体素数据进行聚类分割,并利用三维区域生长获得缺陷的轮廓坐标、质心和体积。通过实验验证,缺陷体积测量的相对误差低于1%,与商用缺陷分析软件3D Slicer的体积测量结果的精度水平一致,克服了受单层切片检测结果影响大、容错率低的问题,提高了三维缺陷检测算法的精度和稳定性,适用于需要对缺陷的位置、体积等三维特征进行定量测量的三维缺陷检测场景。更多还原

【Abstract】 The method of detecting internal defects in precision components based on X-ray images provides key technical support for quality inspection and assessment of precision components in critical fields such as aerospace and nuclear industries.X-ray images include two-dimensional DR(Digital Radiography)images and three-dimensional CT(Computed Tomography)images,two-dimensional DR images are used to qualitatively detect defects,determine the presence of defects and roughly estimate the location and size of the defects,and three-dimensional CT images are used to quantitatively measure the defects.The 3D CT images are used to quantitatively measure the location,contour,volume and other three-dimensional morphological features of the defects.Domestic research on defect detection algorithms based on X-ray images started late,and there is a big gap in detection accuracy and efficiency compared with the foreign advanced level.At present,the internal defect detection methods for precision components based on X-ray images mainly have the following problems that urgently need to be solved:(1)defect detection methods based on artificial feature extraction of two-dimensional DR images are difficult to obtain high discriminative features of defects in low contrast image scenes,resulting in high rates of missed and false detections.(2)Most deep learning defect detection networks based on twodimensional DR images do not consider the imbalance of defect features,have weak anti-interference ability,are easily affected by noise,and are difficult to achieve high detection accuracy.(3)The current defect 3D morphology detection method of 3D CT images based on single-layer slicing is greatly affected by the detection results of single-layer slicing,with low fault tolerance,limited detection accuracy,and low utilization of defect 3D spatial correlation information.Aiming at the above problems,our project researches on the method of internal defect detection in precision components based on X-ray images,and mainly completed the following work:Firstly,aimed at the problem that it is difficult to obtain high discriminative features of defects in manual feature extraction algorithms in defect detection,we designed a defect detection algorithm based on dual threshold reconstruction based on 2D DR images,which uses the ideas of local dynamic threshold segmentation and dual threshold reconstruction to extract high discriminative grayscale features of defects in a local area,in order to reduce the rate of defects leakage and misdetection rate.Through experimental verification,the defect leakage detection rate is 16.9% and the false detection rate is 10.8%,which successfully detects the defects that cannot be detected by artificial feature extraction class methods such as global threshold segmentation and gray gradient extraction,and significantly improves the defect detection accuracy.This method is convenient and fast,with high detection efficiency and low dependence on datasets.It is suitable for two-dimensional defect detection scenarios that require high detection efficiency and are difficult to obtain datasets.Secondly,aiming at the problem that the deep learning defect detection network does not consider the imbalance of defect features,we designed an intelligent defect detection algorithm for 2D DR images based on imbalanced convolutional feature extraction,constructed an imbalanced convolutional feature extraction network,and introduced the distance-attention mechanism to optimize the loss function,to improve the YOLOv8 model.Through experimental validation,the defect missed detection rate of the improved model is 5.5% and the false detection rate is 5.1%.Compared with the original YOLOv8 model,the defect missed detection rate and false detection rate are improved by 4.7% and1.8%,and compared with the defect detection algorithm of 2D DR image based on double threshold reconstruction,the defect missed detection rate and false detection rate are improved by 11.4% and 5.7%,thus,the accuracy of defect detection is greatly improved.Finally,aiming at the problem that the 3D morphology detection method of defects in 3D CT images based on single-layer slices is greatly affected by the detection results of single-layer slices,we designed a 3D morphology detection algorithm of defects based on fuzzy clustering and region growing,which directly performs cluster segmentation on 3D voxel data based on fuzzy clustering ideas,and obtains the contour coordinates,center of mass,and volume of defects by using 3D region growing.Through experimental verification,the relative error of defect volume measurement is lower than 1%,which is consistent with the accuracy level of the volume measurement results of the commercial defect analysis software 3D Slicer,and overcomes the problems of large influence by the single-layer slicing detection results,and improves the accuracy and stability of the defect 3D morphology detection algorithm.更多还原