基于改进马尔科夫随机场与精确高度函数的列车故障图像层次特征匹配

发布时间：2017-12-28 11:23

本文关键词：基于改进马尔科夫随机场与精确高度函数的列车故障图像层次特征匹配　出处：《湖北工业大学》2017年硕士论文　论文类型：学位论文

【摘要】：铁路列车运行故障动态图像检测系统(TFDS)是一套由我国自主研发的基于机器视觉的列车轨边故障图像检测系统。针对TFDS图像颜色单一和背景复杂等特点,利用图像的空间层次性及其故障区域的形状特征,提出基于改进马尔科夫随机场与精确高度函数的列车故障图像层次特征匹配算法,将故障图像识别分成层次模型建立与形状匹配两部分,用以实现列车常见故障的自动检测。从图像像素的空间交互关系入手,采用马尔科夫随机场(MRF),结合图像金字塔与近邻传播理论,提出基于快速自适应MRF的层次分割算法。首先利用小波变换与MRF理论,建立图像多尺度表达的层次模型,接着引入直方图平滑与近邻传播算法自动指定模型层次数,并利用改进的K-means算法实现图像的快速自适应分割,最后依据分割过程中的能量稳定性,将像素标准差微分作为迭代准则,进一步提升运算速度与鲁棒性。在McGill和Weizmann图像数据库上部分图像的测试结果表明,该算法计算速度较快,分割性能好且略优于MRF算法,分割效率较MRF算法提升至少40%。鉴于形状匹配在图像识别中的实用性,在高度函数描述子(HF)的基础上,提出一种精确高度函数特征(EHF)描述算法。首先构造目标形状外轮廓采样点的EHF描述符并进行特征降维,其次利用并行动态规划进行形状匹配,最后引入形状复杂度分析提升匹配效果。基于点的几何特征显著性,提出形状精度理论,进一步分析局部形变与边缘噪声对形状特征描述的影响。在MPEG-7、Swedish Leaf、Tools和ETH-80数据库上进行匹配实验以及在Kimia99数据库上进行抗噪实验,实验结果表明:该算法效率高,匹配时间仅为HF的12.5%,在上述匹配实验中的检索率分别为90.36%、95.07%、94.29%和89.90%,检索性能优于HF和其他重要算法;在抗噪实验中,该算法的抗噪性能也较HF有明显提升。依据TFDS图像的自身特征,结合图像层次特征模型和形状匹配算法,提出的列车故障图像层次特征匹配算法,能实现空气制动系统故障、车轮挡键丢失、高摩合成闸瓦丢失和制动梁安全链脱落四类列检故障的自动识别,该算法缺陷识别率高,鲁棒性好,能有效地应用于TFDS故障图像检测中。
[Abstract]:The dynamic image detection system of railway train running fault (TFDS) is a set of train track fault image detection system based on machine vision, which is developed by our country independently. According to the characteristics of the complex TFDS image of a single color and background, using the shape feature space hierarchy and fault area of the image, put forward improved Markov random field and the exact height function of train fault image level feature matching algorithm based on image recognition, the fault is divided into two parts and establish the hierarchical shape matching model, with automatic detection to achieve common the fault of the train. Starting from the spatial interaction relationship of image pixels, Markov random field (MRF) and image Pyramid and nearest neighbor propagation theory are applied to propose a hierarchical segmentation algorithm based on fast adaptive MRF. Firstly, using wavelet transform and MRF theory, establish the hierarchy model expression of multiscale image, then introduce the histogram smoothing and affinity propagation algorithm to automatically specify the model layer number, segmentation fast adaptive and improved K-means algorithm to realize image segmentation, based on the energy stability in the process of the pixel standard deviation differential as the iterative criterion, further to improve computing speed and robustness. The test results of some images on McGill and Weizmann image databases show that the algorithm is fast and has better segmentation performance and slightly better than MRF algorithm, and the segmentation efficiency is improved by at least 40% compared with MRF algorithm. In view of the practicality of shape matching in image recognition, an exact height function feature (EHF) description algorithm is proposed on the basis of the height function descriptor (HF). First, we construct the EHF descriptor of the target contour and sampling points, and feature reduction. Secondly, we use parallel dynamic programming for shape matching. Finally, we introduce the shape complexity analysis to improve the matching effect. Based on the geometric feature saliency of point, the shape precision theory is proposed to further analyze the influence of local deformation and edge noise on the description of shape feature. In MPEG-7, Swedish, Leaf Tools and ETH-80 database, and the anti noise experiments on the Kimia99 database, the experimental results show that the proposed algorithm is of high efficiency, matching time is only 12.5% of HF, in the matching experiment in retrieval rate were 90.36%, 95.07%, 94.29% and 89.90%, and the retrieval performance is better than HF other important algorithm; the anti noise experiment, the anti noise performance of the algorithm is HF has improved significantly. According to the characteristics of TFDS image, and image level feature model and shape matching algorithm, put forward the train fault image level feature matching algorithm, can realize the air brake system failure, wheel gear key loss, high friction composite brake shoe and brake beam safety chain lost off four class train inspection fault automatic recognition algorithm, the recognition rate of defects high, good robustness, can be effectively applied to the fault detection in TFDS image.
【学位授予单位】：湖北工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U279.3;TP391.41

【相似文献】