【作者】 郭春华；

【导师】 汪同庆；

【作者基本信息】 重庆大学 ， 仪器科学与技术， 2011， 博士

【摘要】 随着我国经济的高速发展以及城市的快速扩张,城市轨道交通已经成为我国城市交通发展的主流,其健康安全状况是保证整个系统安全运行的关键之一。PC梁作为跨座式单轨梁桥系统的重要组成部分,其表面出现裂纹及其它缺陷,对单轨交通的安全运营造成了隐患,目前的人工巡检方式存在效率低下、耗人力、费用高等问题。因此,开展跨座式单轨交通PC轨道梁面裂纹缺陷检测技术的研究,设计有效的检测设备和自动检测算法是当前亟待解决的热点、难点问题。本文从梁面图像检测系统入手,针对自动检测算法中存在的一些难题,从梁面图像的去噪、增强、裂纹提取分割以及裂纹病害的分类、度量等多个方面进行了研究。本文首先给出了PC轨道梁裂纹病害检测的系统框架,采用数据采集系统和离线处理系统两部分组成。数据采集系统从线阵相机、数据采集和保存、照明以及精确定位等几个方面详细讨论了硬件特点以及需要注意的问题,并构建了适用于梁面数据采集的硬件系统。离线处理系统简要介绍了裂纹病害检测模块的各个主要工作流程,设计了基于图像分析的梁面裂纹病害检测策略。针对梁面图像噪声严重、裂纹边缘模糊及信噪比低等问题,本文提出了基于核各向异性扩散的裂纹图像去噪方法。在各向异性扩散的基础上,增加一个边缘增强算子,用于增强微弱的裂纹边缘信息,并且根据噪声均匀分布在多维空间的特点,把低维数据推广到高维空间,结合核方法的优点,在核空间中实现去噪,同时采用平均绝对差值的自动扩散终止规则也提高了核各向异性扩散的效率。实验结果表明,该方法对于PC轨道面线性裂纹图像的去噪有较好的效果,具有较高的信噪比,为后期进行梁面裂纹检测并精确的评价梁面质量打下了坚实基础。梁面图像中的裂纹信息往往较为弱小,与背景的对比度很低,难以直接检测到。基于此,本文提出了基于非下采样Contourlet变换(NSCT)的跨座式单轨梁面裂纹增强新算法。该方法根据NSCT域的不同特性,利用分数阶微分能增强信号的中频成分、非线性保留信号低频的特点,在平滑子带引入分数阶微分增强纹理信息;同时在高频方向子带根据NSCT对方向的敏感性,采用对弱边缘增强、强边缘保留、噪声去除的处理,使得视觉效果和对比度改善指数大大提高,增强效果良好。在裂纹提取分割上,本文提出了两种方法。首先提出基于凸残差的分割方法,由于裂纹缺陷相对背景而言所占比例为零或较小,直方图分布为单峰或接近单峰,该方法把直方图看做平面上的区域,根据待检测图像直方图的凹凸性,计算其凸包函数,并求取凸残差,当凸残差及相应的类间方差整体上达到最大时,所对应的灰度值即为分割阈值。另一种方法是基于曲率的裂纹检测算法,将二维平面图像映射到三维曲面空间,从微分几何出发,通过三维曲面中裂纹的曲率特征,由高斯曲率的极值点和负平均曲率来提取裂纹。对于检测结果中出现的伪裂纹,设计了结合面积、圆形度和长宽比的评判规则,消除过小、过短的缺陷,成功消除了绝大部分伪裂纹。根据不同类型裂纹的几何形态差异,本文通过提取裂纹的投影向量比值、面积以及方向角方差等具有鉴别意义的裂纹模式特征,设计了基于神经网络的分类器实现对裂纹的精确分类。另外在裂纹病害的度量上,采用裂纹缺陷面积、单根裂纹的长度和宽度度量方法,误差在可接受范围内,从而验证了算法的有效性。更多还原

【Abstract】 With the high-speed development of chinese economy and fast expansion of urban, city rail transit system has been the main trend of city traffic mode in China, but its safety is a key problem to the whole transit system. The PC beam is an important component of Straddle Monorail traffic system, the surface of which has been covered with cracks and other flaws, it causes deficiencies to the safety of system. At present, the inspection is made by manual, which exists low efficiency, cost manpower, high cost, and so on. So the crack detection of Straddle Monorail traffic beam surface should be researched. Design efficient equipment and automatic analysis algorithms are the focus and difficulty problems. The paper proposes a crack inspection system for automatic component detection, including inspection equipment and detection methods such as image denoising, enhancement, crack segmentation, flaw classification, flaw measurement, and so on.First, the paper proposes a crack and flaw inspection frame of PC track beam, including data acquisition system and off-line processing system. The data acquisition system consists of line camera selection, data acquisition and save, lighting, precise locating, and other hardware integration problems which should be pain attention, and the hardware platform is built for track beam data collection. The off-line processing system briefly introduces the work flow of different flaw inspection module and the crack detection strategies of track beam base on image analysis.For the problems of track beam image polluted by noise, crack edge blurred and low signal to noise, a new kernel anisotropic diffusion noise removal algorithm is proposed in this paper. On the basis of anisotropic diffusion, an enhance operator which promotes the weak crack edge is added, and according to the characteristics of noise uniformly distributed in the multidimensional space, the low dimensional data is promoted to high dimensional space and denoised in the kernel space, also average absolute difference value of automatic diffusion termination criterion is introduced to enhance the efficiency of diffusion. This method has been applied to noise removal of low signal-noise ratio track beam surface crack, the kernel anisotropic diffusion outperforms other method for denoised result and signal-noise ratio, it thus provide the bases for the behind successful crack detection and precise positioning.The crack is difficult direct detection for it is usually weak, and has low contrast as comparison with background. Then, a new algorithm of crack enhancement of straddle-type monorail track beam surface based on non-subsampled contourlet transformation (NSCT) is presented. It is according to the characteristics of NSCT different domain, the fractional differentiation can enhance the mid-frequency component and retain the low frequency component nonlinearly, then the smooth sub-band texture of NSCT domain is enhanced by the fractional differentiation; at the same time, according to the direction sensitivity of high-frequency sub-band of NSCT domain, the weak edge is enhanced nonlinearly, strong edge is retained, and noise is removed. Experimental results show that the method proposed in this paper has greatly improved visual effects and contrast improve index, and the enhancement effect is good.Two methods for crack extracting have been proposed in this paper. The first one is automatic threshold segmentation approach based on convex residual. The proportion of crack is zero or small comparing to the background, so the histogram distribution is unimodal or close to unimodal. The histogram is regarded as an area of plane, according to the convexity and concavity of histogram of detected image, first compute the convex hull function, then obtain the convex residual which is the difference between convex hull function and the corresponding probability. The maximal value of convex residual join with between-class variance is the threshold value. Another one is based on the curvature of crack. The two dimensional images are mapping to three dimensional surface, starting from the differential geometry, the three dimensional surface curvature characteristics is computed, and the crack is extracted when the gaussian curvature is achieved extreme and the average curvature is negative. There are pseudo cracks appeared in the results, then area, rotundity degree, ratio of length and width are combined to judge between them, they can eliminate defects which are too small or too short, then most of pseudo cracks can be successfully removed.According to the different geometry type of cracks, the ratio of x axis projection and y axis projection of crack, crack area, and angle variance are selected to identify different crack type, and neural network classifier is designed to achieve precise classification of cracks. As well as, for the measure of crack and flaw, crack area, single crack length and width are adopt to measure, and the result error is acceptable, which verifies the validity of the algorithm.更多还原