第一届挑战赛B4-基于图像处理和数据挖掘技术的道路缺陷类型的自动识别

摘要 随着经济的的发展，交通在国民经济和生活中的重要性显著提高。城市道路是城市建设的主要项目之一，工程建设涉及面较广，工程条件较复杂，是由多项目、多工序彼此交错和相互制约所组成的线形工程，影响工程质量的因素较多，施工中不可避免地会出现不同程度的质量问题。为了提高公路使用寿命，公路养护工作也得到越来越多的重视。本文介绍了基于图像处理的路面检测及基于数据挖掘技术的道路缺陷类型自动识别的研究。 首先，通过分析了缺陷路面原始图像，得出了路面图像的特征，为选定图像预处理方法，选择图像特征值和图像分类识别算法建立基础。 其次，研究了路面的预处理问题。为了消除原始图像中的噪声，根据路面图片的特征，本文采用直方图均化、灰度变换方法增强图像，再用加权邻域均值滤波对图像进行平滑处理，通过实验对比几种边缘检测算子的检测效果,证明用 Sobel 算子对图像进行边缘检测的效果最好,同时运用数学形态学运算填充边缘内部的孔洞以及去除图像中孤立和小区域噪声,提取得到裂缝或坑槽目标的二值图像。 最后,在得到目标二值图像后,研究了裂缝目标的特征提取和识别问题。依据分析得到的各类裂缝图像的特点,提取路面裂缝目标的四类特征:第一类是通过垂直投影和水平投影的像素统计图提取裂缝图像的投影特征,第二类是在得到的投影统计图的基础上,根据 Proximity 的算法提取裂缝目标的特征,第三类是利用破损密度因子提取路面裂缝目标的特征，第四类是计算图像的分型维数。 最后基于七个特征向量应用 SVM 算法对路面裂缝图像进行分类识别,通过前人先验的基础上，选取 RBF作为核函数，通过对30 幅图像进行交叉检验实验,通过选取核函数的不同参数进行训练，然后分别进行模型检验 ,通过比较说明本文提供的方法能够比较准确的实现路面缺陷类型的识别。 关键词关键词关键词关键词：：：：数据挖掘；图像处理；路面缺陷类型；模式识别；支持向量机 Pavement Flaw’s Automatic Recognition Based on Image Processing andData Mining Abstract With the rapid development of highway construction and gradual improvement ofroadnetwork construction in China, road maintenance work has been paid more and moreattention.Pavement flaw is the main form of road diseases. It is also an important indicator ofthe roadquality assessment. The traditional manual detection and recognition methods are notable tomeet the requirement of rapid development of highways, so the research of pavementflawautomatic detection and recognition is particularly urgent. Therefore, in this thesis someresearch are done on Pavement Flaw’s Automatic Recognition Based on Image Processingand Data Mining. Firstly, we analyse the characteristics of the sample images,which will be the bases ofimage pre-processing, feature extraction and automatic reconition of the image . Secondly, the research of image pre-processing is made after the characteristics of thepavement flaw image are analyzed. The pavement flaw images which we collected inevitablycontain much noise, which cause many difficulties in classification and recognition ofpavement flaw image. In order to facilitate subsequent operations, the image is enhancedbased on gray transformation and weighted neighborhood average filter. And then,it is provedthat using Sobel operator can get the best result in edge detection with the comparison of theseveral edge detection operators. Based on this, after the holes inside the edge are filled andthe isolated and small regional noises are removed by using mathematical morphologyoperation. Furthermore, the binary flaw image is extracted and the pavement flaw imagesegmentation is completed. Finally, flaw feature extraction and recognition are studied. On the basis of analysisofcharacteristics of various types of pavement flaw characteristics was accomplished, fourkinds of features are extracted from the pavement flaw image. The first is to extract projectionfeatures of pavement flaw image with the vertical projection and horizontal projection of pixelstatistical chart. The second is to extract flaw features based on proximity algorithm aftergetting the projection statistical chart. The third is to extract density factors of features andeffectively reduced noise furthermore. The forthfearure is fractaldimensions of theimages.Then, classification and recognition of the pavement crack image is completed based 第2页，共26页 on SVM algorithm with 7 extracted features. On the basis of former research, we chooseRBF as theBasis Function, then we uses 30 images to do the cross validation, and train themodel by choosing different parameter of the Basis Function.By testing the model, we foundthat the way provided in this thesis can recognize different type of imagesmore precisely. Key words:data mining image processing road suface automatic recognition andclassificationSVM(Support Vector Machine) 目录 1.研究目标研究目标研究目标研究目标............................................................................................... 5 2.分析方法与过程分析方法与过程分析方法与过程分析方法与过程.................................................................................... 5 2.1.总体流程..................................................................................................................... 52.2.具体步骤..................................................................................................................... 62.3.结果分析................................................................................................................... 14 1.挖掘挖掘挖掘挖掘目标目标目标目标 本次建模目标是在缺陷类型的道路图像进行增强去噪等预处理、图像特征值的选择与提取的基础上，利用提取得到的真实数据，采用数据挖掘技术，分析各类道路图像特征值与缺陷类型之间的相互关系，训练自动分类算法，根据分类器的分类结果判断待识别样本属于何种类别的缺陷，从而实现不同道路缺陷类型的自动识别。 2.分析方法与过程分析方法与过程分析方法与过程分析方法与过程 2.1.总体流程总体流程总体流程总体流程 本用例主要包括如下步骤： 步骤一步骤一步骤一步骤一:缺陷道路图像预处理缺陷道路图像预处理缺陷道路图像预处理缺陷道路图像预处理 通过对路面图像的分析原始图像的特征，选择合适的预处理手段，先对图像进行剪切【1】，取出无用部分，再增强图像目标与背景的对比度，使目标边缘平滑，检测目标边缘，最后通过形态学运算（腐蚀、膨胀、开运算、闭运算）减少二值图像的噪声点，为进一步提取图像特征做准备。 步骤二步骤二步骤二步骤二：：：：道路缺陷目标道路缺陷目标道路缺陷目标道路缺陷目标特征特征特征特征提取提取提取提取 根据路面图像预处理后得到的缺陷特征，选择合适的算法，提取出图像的4特征值，作为路面缺陷类型自动识别的基础，分别为（1）基于投影的特征提取（2）基于proximity算法的特征提取（3）基于破损密度因子的路面破损特征提取（4）基于分型的特征提取 步骤三步骤三步骤三步骤三：：：：路面缺陷类型的分类识别路面缺陷类型的分类识别路面缺陷类型的