摘要
道路精准养护关键是精确发现并解决影响车辆驾驶的道路问题,有效缩短道路病害工作闭环时间.针对道路养护中路面病害分割实时性与准确性,提出一种基于深度过参数化卷积的路面病害分割网络.首先,利用Focus模块与2层3×3卷积的切片操作替换了网络特征压缩结构,以减小图像信息丢失.其次,通过替换ResNet50卷积模块中的传统卷积为深度过参数化卷积,提升网络收敛速度,并引入卷积块状注意力机制增强特征提取网络对图像信息的聚焦能力.最后,组合原型网络与预测头网络分别生成的原型掩膜与预测框掩膜系数,完成路面病害分割.路面病害分割实验在公开数据集和自制数据集下进行,分割平均精度AP_(all)分别为21.59%和31.43%,分割速度分别为31.33帧/s和30.52帧/s.实验结果表明改进后模型能够实现路面病害分割的实时性与高精准性.
The key to accurate road maintenance is to accurately find and solve the road problems affecting vehicle driving,and effectively shorten the closed-loop time of road diseases.Aiming at the real-time and accuracy of pavement disease segmentation in road maintenance,a pavement disease segmentation network based on depthwise over-parameterized convolutional layer is proposed.Firstly,slice with focus module and two-tier 3×3 convolution replaces the network feature compression structure to reduce the loss of image information.Secondly,by replacing the traditional convolution block in ResNet50 with depthwise over-parameterized convolutional layer,the convergence speed of the network is improved,and the convolution block attention mechanism is introduced to enhance the focusing ability of feature extraction network on image information.Finally,the prototype mask and prediction mask coefficients generated by Prototypical Networks and Prediction Head Network are combined to complete the pavement disease segmentation.The pavement disease segmentation experiment is carried out under the public data set and the self-made data set,and the segmentation average accuracy AP_(all) is 21.59%and 31.43%respectively,and the segmentation speed is 31.33 frames/s and 30.52 frames/s respectively.The experimental results show that the improved model can achieve real-time and high accuracy of pavement disease segmentation.
作者
刘玉文
黄友锐
韩涛
LIU Yuwen;HUANG Yourui;HAN Tao(School of Artificial Intelligence,Anhui University of Science and Technology,Huainan 232001,China;School of Electrical and Information Engineering,Anhui University of Science and Technology,Huainan 232001,China;School of Electrical and Optoelectronic Engineering,West Anhui University,Lu’an 237012,China)
出处
《湖北民族大学学报（自然科学版）》
CAS
2022年第4期437-444,共8页
Journal of Hubei Minzu University:Natural Science Edition
基金
国家自然科学基金项目(61772033)
安徽省高校协同创新项目(GXXT-2019-048,GXXT-2020-54).
关键词
路面病害图像
实例分割
深度过参数化卷积
注意力机制
pavement disease image
instance segmentation
depthwise over-parameterized convolutional layer
attention mechanism
