基于改进两步式卷积神经网络的沥青路面裂缝识别方法

A Method for Recognizing Asphalt Pavement Cracks Based on Modified Two-step Convolutional Neural Network

为实现沥青路面裂缝长度和面积等几何信息的智能提取,提出了一种基于改进两步式卷积神经网络的沥青路面裂缝几何信息提取法。在完成数据的采集后,第1步主要通过人工筛选出失真图像、无病害图像和可能存在病害图像各600张,训练基于卷积注意力模块改进的图像分类算法ResNet50,以此作为本模型的清洗算法,完成对410 000张原始路面图像的清洗工作,筛除其中的失真图像和无病害图像,构建裂缝病害图像数据集;第2步基于卷积注意力模块对图像语义分割算法U-Net进行改进,使用上一步筛选的数据进行训练与测试,训练集与测试集的样本比例为10∶1,以此实现裂缝图像的分割,并对裂缝的长度和面积进行提取。试验结果表明,本研究提出的改进ResNet50算法对于总体样本清洗结果的精确率、召回率和F_(1)值均已超过95%,其中F_(1)值已经达到了96.8%;两步式沥青路面裂缝几何信息提取法的均交并比为0.496 7,其中横向裂缝、纵向裂缝、龟裂、块状裂缝的交并比分别为0.495 1,0.546 7,0.608 5和0.336 6;在裂缝长度信息提取上,横向裂缝的误差为12.18%,纵向裂缝的误差仅为3.88%;在裂缝面积的信息提取中,横向裂缝、纵向裂缝、龟裂、块状裂缝的误差分别为0.58%,12.14%,12.27%,11.44%。

A modified two-step convolutional neural network(CNN) method was investigated to analyze pavement crack length and area in the study. Firstly, an image classification algorithm called ResNet50 was modified by a Convolutional Block Attention Module(CBAM) and then used to eliminate distortion and non-defect images from 410 000 raw samples. In the second step, an image segmentation algorithm called U-Net, also modified by CBAM, was used to segment cracks from non-crack regions. Length and area of pavement cracks were extracted. Results show that precision, recall, and F_(1)-score of the modified ResNet50 exceeds 95%, and F_(1)-score reaches 96.8%;Mean Intersection over Union(MIoU) of the two-step method is 0.496 7;Intersection over Union(IoU) of transverse crack, longitudinal crack, craze crack, and block crack are 0.495 1,0.546 7,0.608 5,0.336 6 separately;prediction error of length from transverse crack, and longitudinal crack is 12.18% and 3.88% separately. Prediction error of area from transverse crack, longitudinal crack, craze crack, and block crack are 0.58%,12.14%,12.27%,11.44% separately.