基于改进Faster R-CNN的路面灌封裂缝检测方法

Pavement Sealed Crack Detection Method Based on Improved Faster R-CNN

路面灌封裂缝对路面使用寿命的影响较为突出,为了解决目前灌封裂缝检测技术匮乏的问题,文中提出了一种基于改进Faster R-CNN的路面灌封裂缝检测方法。首先,建立灌封裂缝图像集,对采集到的图像进行增广处理,构建路面灌封裂缝标注样本数据集,并将图像集按6∶2∶2的比例划分为训练集、验证集和测试集;接着,采用Faster R-CNN模型对灌封裂缝进行检测,针对Faster R-CNN检测灌封裂缝存在漏检、定位效果不够理想的问题,文中分别将VGG16、ZFNet和Resnet50网络的特征提取层与Faster R-CNN模型进行结合,结果表明,VGG16和Faster R-CNN结合的模型检测精度最高,达到0.9031;然后,通过增加灌封裂缝候选框宽高比的方法继续改进模型,检测精度达到0.9073,且原先被漏检的目标能被检测出来;最后,对改进Faster R-CNN与YOLOv2模型的检测精度及定位效果进行对比,结果表明,文中提出的改进Faster R-CNN能够明显提高对灌封裂缝的检测准确率和定位精度。

Pavement sealed cracks have significant impact on service life of pavement.A new method for pavement sealed crack detection based on improved faster R-CNN was proposed,aiming at solving the current lack of sealed crack detection technology.Firstly,the marked sample data set of pavement sealed crack was constructed based on the augmented sealed crack image set.Then they were divided into training set,verification set and test set accor-ding to the ratio of 6∶2∶2.Next,faster R-CNN model was employed in sealed cracks detection.Given that the faster R-CNN model has the demerits of miss detection and inaccurate positioning of sealed cracks,it was combined the feature extraction layers of VGG16,ZFNet and ResNet 50 networks.The results show that the detection accuracy of the VGG16 and faster R-CNN combination models can reach 0.9031,which is the highest.Then,further improvement was made by increasing the aspect ratio of the anchor of the sealed crack.The improved detection accuracy reaches 0.9073 and the original miss detection target can also be detected.Finally,detection and positioning accuracy between improved faster R-CNN and YOLOv2 model was compared.The result shows that improved faster R-CNN model can significantly enhance both detection and positioning accuracy.