摘要
为在现有车载计算平台上实现实时路面识别,基于ResNet50网络提出了一种轻量化网络DSC-ResNet50.对于ResNet50,将3×3的深度卷积层分别替换残差单元中的3×3普通卷积层,并从理论上推导了轻量化率改进能够减低约50%的参数量.利用公共数据集CIFAR-10对提出的DSC-ResNet50网络予以验证,结果显示参数量和计算量分别减少了47.96%和44.49%.利用DSC-ResNet50网络在开源路面图像分类数据集(RSCD)进行了路面类型识别研究,结果表明:DSC-ResNet50的参数量和计算量较传统ResNet50网络分别减少了47.87%和44.47%,同时准确率提高了1.07%,验证了所提出方法的有效性.
In order to realize real-time road identification on the existing vehicle computing platform,a lightweight network DSC-ResNet50 was proposed based on RESNET 50 network.For ResNet50,the depth convolution layer of 3×3 was used to replace the ordinary convolution layer of 3×3 in the residual unit,and it was theoretically deduced that the improvement of lightweight ratio can reduce the parameter quantity by about 50%.The public data set CIFAR-10 was used to verify the proposed DSC-ResNet50 network.The results show that the parameters and calculation amount are reduced by 47.96% and 44.49% respectively.Using DSC-ResNet50 network,the pavement type recognition was studied in the open source pavement image classification data set(RSCD).The results show that the parameters and calculation amount of DSC-ResNet50 are reduced by 47.87% and 44.47% respectively compared with the traditional ResNet50 network,and the accuracy rate is improved by 1.07%,which verifies the effectiveness of the proposed method.
作者
骆熠
文国军
林科
LUO Yi;WEN Guojun;LIN Ke(School of Mechanical and Electronic Information,China University of Geosciences(Wuhan),Wuhan 430074,China;Hubei Intelligent Geological Equipment Engineering Technology Research Center,Wuhan 430074,China;Key Laboratory of Geological Survey and Evaluation of Ministry of Education,China University of Geosciences(Wuhan),Wuhan 430074,China)
出处
《武汉理工大学学报(交通科学与工程版)》
2025年第4期871-876,共6页
Journal of Wuhan University of Technology(Transportation Science & Engineering)
关键词
路面识别
残差网络
深度卷积
轻量化改进
RSCD数据集
road surface recognition
residual network
depthwise convolution
lightweight improvement
RSCD dataset
