基于改进距离正则化水平集的钢轨表面缺陷提取方法

An Extraction Method of Rail Surface Defects Based on Improved Distance Regularized Level Set

针对现有钢轨表面缺陷检测方法复杂度过高且精确率较低,提出基于Sigmoid对比度拉伸的距离正则化水平集算法。通过对数变换和均值操作对钢轨图像进行预处理;根据钢轨表面图像梯度变化,以Sigmoid对比度拉伸方法替换距离正则化水平集中的停止函数;根据初始轮廓曲线演化方向的特点,通过新的自适应速度函数替换距离正则化水平集中的演化速度常量;以曲线演化的最终轮廓提取缺陷区域。仿真实验显示,新算法在钢轨表面缺陷数据集中的精确率为99.07%,豪斯多夫距离平均值为6.64,检测缺陷用时为16.46s。与DRLSE、DCNNs和CTFM算法相比,改进的距离正则化水平集算法提升了检测的精确率和检测速度,对噪声具有一定的鲁棒性并适用小范围缺陷检测。

Detection methods of rail surface defects concerning high complexity and low accuracy. An algorithm of distance regularized level set based on Sigmoid contrast stretching method is proposed. Firstly, preprocess images of rails through logarithm transform and mean operation. Then, according to gradient change of rail surface image, use Sigmoid contrast stretching method to replace edge stopping function of the distance regularized level set. Considering moving tendency and features of initial contour curve, a new adaptive velocity function is used to replace the evolution velocity constant of the distance regularized level set. Finally, defect areas are located by analyzing a final contour of curve evolution. The simulation shows the accuracy rate of the algorithm in data set of rail surface defects is 99.07%;the average value of Hausdorff distance is 6.64;and the detection defect time is 6.64 s. Comparing with DRLSE, DCNNs, and CTFM, the advanced distance regularized level set improves detection precision and testing speed. It has a more preferable robustness to noise and is suitable for small-scale defect detection.