基于深度卷积网络的路面裂缝分割方法 李伟,申浩',马志丹’,孙朝云',呼延菊2 (a.长安大学信息工程学院,陕西西安710064:b.加拿大滑铁卢大学,加拿大滑铁卢N2L3G1) 摘要:路面裂缝作为路面常见病害之一,是公路养护管理工作的重点。将裂缝从路面 背景中精确的分割是裂缝检测和修补的基础。传统的裂缝分割方法易受到光照强度 和路面噪声的影响,分割效果有待提高。本文提出了一种基于深度卷积网络的路面 裂缝分割方法。该方法采用深度卷积网络对路面裂缝进行特征学习和提取,同时结 合多尺寸特征图进行裂缝的分割。通过采用深度卷积网络,可以克服多种因素对裂 缝分割的干扰。通过过联合多尺寸特征图对裂缝进行分割,可以提高神经网络分割 的准确性。本文冋时将该方法与阈值分割方法、F娳N进行了对比。结果表明,本文 提出的裂缝分割方法明显优于其他方法。 关镳词:路面裂缝;深度卷积网络;裂缝分割 中图分类号:U238 An Innovation of pavement Crack segmentation Based on deep convolutional Li Wei!, Shen Hao, Ma Zhi-dan, Sun Zhao-yun, Huyan ur (1. School of Information Engineering, Chang an University, Xian 710064; 2. University of Waterloo. Canada. Waterloo. Canada N2L3G1) Abstract: Pavement crack as one of the common diseases is the key point of highway maintenance and management. Precise segmentation of crack from the background is the fundament of crack detection and mending. Traditional crack segmentation methods are easily affected by light intensity and pavement noise, and the accuracy of segmentation needs to be improved. An innovation of pavement crack segmentation based on deep convolutional network is put forward in this paper. The deep convolution network is used to learn and extract the characteristics of pavement cracks, and multi-dimensional feature maps are combined for the segmentation of cracks. By using deep convolution network, the interference of many factors to crack segmentation can be settled. The accuracy of neural network segmentation can be improved by combining multi-dimensional feature maps. This method is compared with threshold method and FCN, the results show that the proposed method is 基金项目:陕西省自然科学基金-重大基础研究项目(2017ZDJC-23) 通讯作者:申浩,男,长安大学硕士研究生。 Email: shen8927 foxmail. com基于深度卷积网络的路面裂缝分割方法 李伟 1 ，申浩 1* ，马志丹 1 ，孙朝云 1 ，呼延菊 2 （a.长安大学信息工程学院，陕西 西安 710064； b.加拿大滑铁卢大学,加拿大 滑铁卢 N2L3G1） 摘 要： 路面裂缝作为路面常见病害之一，是公路养护管理工作的重点。将裂缝从路面 背景中精确的分割是裂缝检测和修补的基础。传统的裂缝分割方法易受到光照强度 和路面噪声的影响，分割效果有待提高。本文提出了一种基于深度卷积网络的路面 裂缝分割方法。该方法采用深度卷积网络对路面裂缝进行特征学习和提取，同时结 合多尺寸特征图进行裂缝的分割。通过采用深度卷积网络，可以克服多种因素对裂 缝分割的干扰。通过过联合多尺寸特征图对裂缝进行分割，可以提高神经网络分割 的准确性。本文同时将该方法与阈值分割方法、FCN 进行了对比。结果表明，本文 提出的裂缝分割方法明显优于其他方法。 关键词： 路面裂缝；深度卷积网络；裂缝分割 中图分类号：U238 An Innovation of Pavement Crack Segmentation Based on Deep Convolutional Network Li Wei1 , Shen Hao1 , Ma Zhi-dan1 , Sun Zhao-yun1 , Huyan Ju2 （1. School of Information Engineering, Chang’an University, Xi’an 710064; 2. University of Waterloo, Canada, Waterloo, Canada N2L3G1） Abstract: Pavement crack as one of the common diseases is the key point of highway maintenance and management. Precise segmentation of crack from the background is the fundament of crack detection and mending. Traditional crack segmentation methods are easily affected by light intensity and pavement noise, and the accuracy of segmentation needs to be improved. An innovation of pavement crack segmentation based on deep convolutional network is put forward in this paper. The deep convolution network is used to learn and extract the characteristics of pavement cracks, and multi-dimensional feature maps are combined for the segmentation of cracks. By using deep convolution network, the interference of many factors to crack segmentation can be settled. The accuracy of neural network segmentation can be improved by combining multi-dimensional feature maps. This method is compared with threshold method and FCN, the results show that the proposed method is 基金项目：陕西省自然科学基金-重大基础研究项目（2017ZDJC-23） 通讯作者：申浩，男，长安大学硕士研究生。Email: shen8927@foxmail.com