基于SVM和改进区域生长法的桥梁裂缝分割算法

A Bridge Crack Segmentation Algorithm Based on SVM and Improved Region Growing Method

为了解决复杂背景下裂缝分割算法泛化能力差,图像分割精度不够高,裂缝边缘分割不够精确,裂纹分割结果连续性较差的问题,提出了一种基于改进区域生长法桥梁裂缝分割算法。首先对裂纹图像全局灰度化处理,对灰度化后裂纹图像进行双重滤波,利用最大类间方差法(Otsu)对经滤波优化的裂缝图像进行粗分割。接着,通过分割得到了裂缝连通域,对其采用最小外接矩形长宽比、面积、圆形度、标准差特征进行了支持向量机训练,利用训练结果对连通域进行分类,此步骤可以较好地减少误判。然后,提出了一种自动选择高质量起始种子点集的方法,并对种子点集中的点进行验证,以种子点集中的点作为起始点开始区域生长,很好地避免了人工选择种子点效率低下、种子点质量不稳定的问题。最后,将经过区域生长后的裂缝图像进行形态学开运算处理去除毛刺,再进行形态学闭运算去除空洞、过滤掉孤立区域,根据裂缝连通域间方向一致性和相对位置信息对裂缝进行连接。以背景包括渗水、混凝土砂浆黏结等干扰的裂缝图像作为试验图像进行试验。结果表明:本研究算法分割表现良好,算法波动较小,泛化好,准确率均值达到98.798%,精确值均值达到85.686%,召回值均值达到88.579%,F1值均值达到86.572%。

In order to solve the problems of poor generalization ability,insufficient image segmentation accuracy,inaccurate crack edge segmentation and poor continuity of crack segmentation result under complex background,a bridge crack segmentation algorithm based on improved region growing method is proposed.First,the crack image is gray-processed globally,the gray-processed crack image is double(Bilateral Frangi)filtered,the filtered and optimized crack image is coarsely segmented by the maximum inter-class variance(Otsu)method.Then,the crack connected domain is obtained by segmentation,the minimum outer rectangle aspect ratio,area,circularity and standard deviation are used for SVM training on it.The connected domain is classified by using the training result,this step can better reduce misjudgment.Afterwards,a method for automatic selecting high-quality initial seed point set is proposed,and the points in the seed point set are verified.Taking the points in the seed point set as the starting points to start the region growth,which avoids the problems of low efficiency of and unstable quality of manual selected seed points.Finally,the crack image after regional growth is processed by morphological open operation to remove burrs.The morphological closing operation is also carried out to remove cavities and filter out isolated areas,and cracks are connected according to direction consistency and relative position information between connected areas of cracks.The crack image with the background including water seepage and concrete mortar bonding is used for test.The result shows that the segmentation performance of the proposed algorithm is good,the algorithm has little fluctuation and good generalization,the average accuracy rate reached 98.798%,the average precision value reached 85.686%,the average recall value reached 88.579%,and the average F1 value reached 86.572%.