基于深度卷积神经网络融合模型的路面裂缝识别方法

Pavement crack identification method based on deep convolutional neural network fusion model

现有的路面裂缝识别方法大多仍局限于基于主动特征提取的处理技术,对路面图像来源有专一性要求,算法不具备泛化能力,现有的基于神经网络识别算法对设备有特定要求,且路面裂缝的定位准确性不高。为此,提出基于深度卷积神经网络融合模型的路面裂缝识别方法。首先,应用多目标SSD卷积神经网络模型对路面裂缝进行分类检测,然后使用深度残差网络对SSD模型的特征提取结构进行改进,并根据损失函数的收敛程度对模型中的超参数进行优化,提高路面裂缝分类和定位的准确率;其次,针对裂缝分类检测模型对路面裂缝定位存在的偏差,提出基于U-Net模型的路面裂缝分割方法,并改进模型的特征提取网络,提高裂缝分割精度,实现精确的裂缝分割;最后,将裂缝分类检测模型与分割模型进行融合,加载2个模型并导入上述训练得到最优权重,根据裂缝分类网络判断路面图像有无裂缝,若存在裂缝则给出具体类别和置信度,并将这些信息和原始裂缝图像输入U-Net分割网络,根据分割结果计算线性裂缝的长度、宽度及网状裂缝的面积。试验结果表明:给出的路面裂缝识别方法对于横向裂缝、纵向裂缝和网状裂缝的识别精度分别为86.6%、87.2%和85.3%;该方法不仅能够给出路面裂缝的类别信息,还可以给出路面裂缝的精确定位和几何参数信息,可直接用于路面状况评价。

Aimed at the existing methods of using deep learning technology for road surface crack recognition,most of the methods were still limited to the processing technology based on active feature extraction,and the source of the road image was targeted,resulting in the algorithm didn’t have the ability to generalize the identification algorithm of the network still has specific requirements for the equipment,and the problem of the accuracy of the location of the road cracks was not high.A method based on deep convolutional neural network fusion model for pavement crack identification was proposed.Firstly,the multi-target SSD convolutional neural network model was applied to classify and detect pavement cracks.Then the depth residual network was used to improve the feature extraction structure of the SSD model,and the hyper parameters in the model were optimized according to the convergence degree of the loss function,which improves the accuracy of classification and location of pavement cracks.Secondly,the classification of cracks detection model was deviations from the pavement crack location.The U-Net model based pavement crack segmentation method was proposed,and the feature extraction network of the model was improved.The crack segmentation accuracy was improved and the precise crack segmentation was realized.Finally,the crack was obtained.The classification detection model was merged with the segmentation model,two models were loaded,and the optimal weights obtained by the above training were imported,and the road surface image was judged according to the crack classification network to determine whether there was a crack,and if there was a crack,a specific category and confidence were given,and these were the information and the original crack image were input into the U-Net segmentation network,the length and width of the linear crack and the area of the mesh crack were calculated according to the segmentation result.The results show that the recognition accuracy of the pavement crack identification method for transverse cracks,longitudinal cracks and map cracks are 86.6%,87.2%and 85.3%,respectively.This method can not only give the category information of pavement cracks,but also give accurate positioning and geometric parameter information,and can be directly used for pavement condition evaluation.5 tabs,19 figs,25 refs.