基于SVM-Adaboost裂缝图像分类方法研究

发布时间：2018-08-19 10:59

【摘要】：不同的裂缝类型关系到不同养护策略。SVM在解决小样本、非线性、高维度问题时具有较大优势,通过采用不同的SVM分类方法和核函数对常用的数据集中的样本进行分类结果对比,选取了RBF核函数和One-against-All的分类方法。但分类结果仍然满足不了路面养护要求。由于Adaboost选择不同的样本进行训练,改变了训练样本的数据分布。每次迭代都会计算得到一个分类效果最佳的弱分类器及其所在总体分类器中的权重。随着迭代次数的增加,最终由弱分类器迭代生成的强分类器的分类误差最小。提出了SVM-Adaboost分类器动态的对SVM参数进行优化。试验结果表明,应用基于SVM-Adaboost的裂缝分类算法对指定样本进行测试,横向裂缝准确率87.48%,纵向裂缝准确率95.37%,网状裂缝准确率94.9%,块状裂缝准确率89.7%。该方法可以提高组合分类器整体的分类精度。
[Abstract]:Different fracture types are related to different maintenance strategies. SVM has great advantages in solving small sample, nonlinear and high dimensional problems. Different SVM classification methods and kernel functions are used to compare the classification results of samples in common data sets. RBF kernel function and One-against-All classification method are selected. However, the classification results still can not meet the requirements of pavement maintenance. Because Adaboost selects different samples for training, the data distribution of training samples is changed. In each iteration, a weak classifier with the best classification effect and its weight in the overall classifier are obtained. With the number of iterations increasing, the classification error of the strong classifier generated by the weak classifier is minimum. The SVM-Adaboost classifier is proposed to dynamically optimize the SVM parameters. The test results show that the accuracy rate of transverse crack is 87.48, longitudinal crack is 95.377.37, net crack is 94.9, block crack is 89.7. This method can improve the whole classification accuracy of the combined classifier.
【作者单位】：交通运输部科学研究院;交通运输部公路科学研究院;
【基金】：交通运输部建设科技项目(2014318J21060)
【分类号】：TP391.41;U418.66

【相似文献】