基于多特征融合的混凝土结构表面病害图像分类算法

Multi-feature fusion based classification algorithm of surface disease image of concrete structure

为提升混凝土表面开裂、露筋锈蚀和损伤3类病害图像分类效率与准确性,减少人工成本,提出了基于多特征融合的混凝土结构表面病害图像分类算法。该算法通过提取混凝土表面病害图像的纹理特征、灰度直方图特征、颜色特征,以支持向量机(SVM)为分类器,分别训练3类特征的病害图像分类模型,采用特征权重算法估计各个特征的权重系数,借助分类模型与权重系数估计病害图像类别划分的可信度值,根据最小误差原则将病害图像判定为可信度值最大的类别。设计了7组覆盖上述图像特征的特征融合方案,以2400张病害图片为样本,训练了3种单一特征和4种多特征病害图像分类模型,并测试算法的准确性。结果表明:总体上,多特征融合分类模型对于混凝土表面病害图像分类准确率高于单一特征分类模型;基于多项式核函数,灰度+颜色+纹理特征融合分类模型分类效果最佳,平均分类准确率达到84%,较单一特征分类模型提升了7%;依赖于混凝土病害图像多特征的综合信息,灰度+颜色+纹理特征融合分类模型可将混凝土表面锈蚀和缺损的分类准确率提升至88%,多特征融合分类模型对于病害的分类判别稳定性显著优于单一特征分类模型。该研究可为混凝土表面病害图像分类提供有效方法,提升病害分类效率与准确性。

To improve the accuracy and efficiency of disease image classification at concrete surface diseases of cracking,bare bar corrosion and defect,and reduce labor cost,one type of disease image classification algorithm of concrete structure surface based on multi-feature fusion was proposed.For this algorithm,texture features,gray histogram features,and color features of disease image were extracted and support vector machine(SVM)was employed as classifier.Three disease classification models were developed,and the three types of image features were taken as training samples,respectively.The algorithm of term weighting was adopted to evaluate the weight of each type of image features,then the classification models and the weights were combined to calculate the reliability value of disease image classification.According to the minimal error rule,the image was classified as the disease type associated with maximum reliability value.Seven feature fusion alternatives were generated by combination of the three types of features,with which three single feature algorithm models and four multi-feature fusion algorithm models were developed with 2400 image training samples for testing the accuracy of the proposed algorithm.The results show that the classification accuracy of multi-feature fusion algorithm model is higher than that of single feature algorithm model.The optimal effect is obtained with the algorithm model fused all the three types of features using polynomial kernel function,with which the average classification accuracy reached 84%,and about 7%higher than single feature algorithm models.Depending on the comprehensive information provided by multi-feature fusion of disease image,the classification algorithm model fused with texture features,gray histogram features,and color features raised the classification accuracy of bare bar corrosion and defect diseases to 88%.Furthermore,multi-feature fusion algorithm model appears a better stability on disease classification when comparing with single feature algorithm model.The research is an effective method for the classification of concrete surface disease,can enhance the efficiency of disease image processing and classification accuracy.2 tabs,6 figs,28 refs.