【作者】 李鑫；

【导师】 余剑武； 陈健美；

【作者基本信息】 湖南大学 ， 机械工程（专业学位）， 2019， 硕士

【摘要】 拉索是斜拉桥的重要构件,对桥梁安全至关重要。桥梁拉索外侧通常有高聚乙烯护套,护套在光照、风、水汽等作用下会产生裂纹、孔洞等缺陷,这些缺陷会对桥梁安全造成潜在威胁。对护套的定期检测和养护能够保证拉索的使用安全。本文对基于机器视觉的缺陷检测方法进行了研究,形成了一套较为完整的检测方案,有望在今后代替人工完成护套缺陷检测工作。本文的主要工作内容有:1)设计了表面缺陷视觉检测方案,其中图像采集模块包括CMOS工业相机、8mm定焦镜头、环形LED工业光源,用于采集护套表面图像。线上算法主要完成图像灰度化、图像去噪、图像分割与缺陷判别,并保存缺陷图像;线下算法完成图像拼接、特征提取、特征选择和图像分类任务。2)研究了图像预处理与图像分割算法。使用YUV颜色空间对采集的图像进行灰度化,用自适应中值滤波算法进行去噪处理,以保留原始图像信息。用sobel算子配合阈值分割、形态学处理等方法完成了缺陷部分的图像分割,最后使用面积和长轴长度判别是否为缺陷图像,长轴角度用PCA算法计算获得。3)对于缺陷可能存在于相邻相机交界处的情况,研究了图橡拼接算法。在特征点选取阶段,使用特征稳定、信息丰富的SIFT算法,并使用FAST特征点提取算法替代SIFT的特征点选取部分,实现了计算效率的大幅提升。经过筛选去除错误匹配点对后,使用RANSAC方法计算最佳的转换矩阵。最后选用灰度加权平均法对两张图像进行融合,得到最终的拼接结果。4)为了获得更有效的信息,对检测到的缺陷图像进行分类,对缺陷分类及相关算法进行了研究。从采集到的图像中提取了16种备选特征,在此基础上使用特征选择算法选出其中10种可分性较好的特征作为分类特征,该过程能够减少信息维度,去除冗余信息,从而提高分类算法的效率和准确率。分类使用SVM分类器完成,对于影响分类器性能的参数c和?,使用狼群算法进行优化,最终的分类器达到97.5%的分类准确度。本文对桥梁拉索表面缺陷检测中涉及的关键算法进行了研究与实验,其结果表明本文中的算法能够有效地去除图像中原有的干扰信息,保留完整的缺陷图像,并提取有效的缺陷特征为分类奠定基础,最后经过智能算法优化提高了缺陷分类的准确度。此外,本文算法的实时性较好,能够满足实际检测的需求。更多还原

【Abstract】 Cables are an important component of cable-stayed bridges and are critical to bridge safety.The outer side of the bridge cable usually has a high-polyethylene sheath,and the sheath will have defects such as cracks and holes under the action of light,wind and water vapor.These deficiencies pose a potential threat to bridge safety Regular inspection and maintenance of the sheath ensures safe use of the cable.In this paper,the machine vision-based defect detection method is studied,and a relatively complete detection scheme is formed.It is expected to replace the manual inspection of sheath defects in the futureThe main work of this paper is:1)A surface defect visual inspection scheme is designed,in which the image acquisition module includes a CMOS industrial camera,an 8mm fixed focus lens,and a ring LED industrial light source for collecting the surface image of the sheath.The online algorithm mainly completes image graying,image denoising,image segmentation,defect discrimination and save defective images.The offline algorithm completes image stitching,feature extraction,feature selection and image classification tasks2)Image preprocessing and image segmentation algorithms are studied.The acquired image is grayed out using the YUV color space,and the adaptive median filtering algorithm is used for denoising to preserve the original image information The image segmentation of the defect is completed by the method of sobel operator combined with threshold segmentation and morphological processing.Finally,the area and the length of the long axis are used to determine whether it is a defect image,and the long axis angle is calculated by PCA algorithm3)For the case where defects may exist at the junction of adjacent cameras,the image stitching algorithm is studied.In the feature point selection stage,the feature-stable and information-rich SIFT algorithm is used,and the FAST feature point extraction algorithm is used to replace the feature point selection part of SIFT,which greatly improves the computational efficiency.After screening to remove the pair of mismatched points,the RANSAC method is used to calculate the optimal transformation matrix.Finally,the two images are merged by using the gray weighted average method to obtain the final stitching result4)In order to obtain more effective information,the detected defect images are classified,The classification of defect and related algorithms are studied.Sixteen candidate features are extracted from the acquired images.Based on this,feature selection algorithms are used to select 10 features with good separability as classification features.This process can reduce the information dimension and remove redundant information.Thereby improving the efficiency and accuracy of the classification algorithm.Classification is done using the SVM classifier,and the parameters c and o-that affect the performance of the classifier are optimized using the wolf group algorithm,and the final classifier achieves 97.5%classification accuracy.In this paper,the key algorithms involved in the detection of bridge cable surface defects are studied and experiment.The results show that the algorithm in this paper can effectively remove the original interference information in the image and keep it intact.The defect image and the extraction of effective defect features lay the foundation for classification.Finally,the optimization of the intelligent algorithm improves the accuracy of the classification of defect.In addition,the algorithm in this paper has good real-time performance and can meet the needs of actual detection.更多还原