基于改进掩膜循环卷积神经网络的沥青路面积水分布检测  

Detection of Water Distribution in Asphalt Road Based on the Improved Mask R-CNN

作　　者：杨炜[1] 常晓博 刘哲[1] 屈晓磊 Yang Wei;Chang Xiaobo;Liu Zhe;Qu Xiaolei(Chang’an University,Xian 710061;Beihang University,Beijing 100191)

机构地区：[1]长安大学,西安710061 [2]北京航空航天大学,北京100191

出　　处：《汽车技术》2022年第8期47-53,共7页Automobile Technology

基　　金：国家自然科学基金项目(U1864204)。

摘　　要：针对传统的掩膜循环卷积神经网络(Mask R-CNN)在检测和分割积水区域时过度重视对区域的分割,忽视了边界分割的问题,在传统Mask R-CNN模型中添加以拉普拉斯(Laplacian)算子为基准构建的边界加权损失函数,得到改进的Mask R-CNN模型,同时以残差网络(ResNet)作为特征提取网络,采用迁移学习的方法,使用自建的1753张路面积水图像进行训练,以查准率与查全率的加权调和平均值F1和雅卡尔(Jaccard)系数作为评价指标。结果表明,改进的模型F1和雅卡尔系数分别为89.23%和81.46%,在不同积水区域、不同环境下,所提出的方法识别准确率和分割精度较高,能够为无人驾驶车辆提供路面积水信息的技术支持。For the traditional Mask R-CNN model which pays too much attention to the detection and segmentation of ponding areas and ignores the segmentation of boundaries,a boundary weighted loss function constructed by Laplacian operator focusing on boundary segmentation is added to the traditional Mask R-CNN model to obtain an improved Mask R CNN model,and ResNet101 is used as the feature extraction network.Using the method of transfer learning,1753 road area ponding water images built by the experiment are used for training.F1 and Jaccard value are used as evaluation indexes in this paper.Research shows that the F1 and Jaccard value of the improved model are 89.34%and 81.46%respectively.The proposed method has higher recognition accuracy and segmentation accuracy in different ponding areas and different environments,and provides road area water information for driverless vehicles.

关 键 词：沥青路面 积水检测 掩膜循环卷积神经网络 实例分割 

分 类 号：U461.91[机械工程—车辆工程]