基于改进SSD算法的路面破损检测  

Pavement damage detection based on improved SSD

作　　者：韦正璐 王家晨 刘庆华 WEI Zhenglu;WANG Jiacheng;LIU Qinghua(Computer College,Jiangsu Science and Technology University,Zhenjiang 212000,China)

机构地区：[1]江苏科技大学计算机学院,江苏镇江212000

出　　处：《电子设计工程》2023年第3期63-68,共6页Electronic Design Engineering

基　　金：国家自然科学基金项目(51008143);江苏省六大高峰人才项目(XYDXX-117)。

摘　　要：针对路面破损图像检测精度低、速度慢等问题,提出一种基于改进SSD的路面破损检测算法。该算法利用经过改进的Inception模块,使其替代SSD网络中的Conv4_3、FC7、Conv8、Conv9和Conv10层;针对路面裂缝、凹陷图像检测识别率低的问题,使用空洞卷积,在不丢失分辨率的情况下扩大感受野,提高浅层特征图对目标物体的特征提取能力。为使网络可以根据输入信息的多个尺度自适应调节接受域大小,在特征提取层引入SKNet。实验结果表明,改进后的方法在自制的路面特征数据集上的平均精度为89.52%,检测速度达到56.23 FPS,较改进前的算法在精度和检测速度方面分别提高了6.07%和13.21 FPS。基于SSD的改进算法不仅能有效识别路面破损特征,而且能有效提高识别速度并提升识别效率。Aiming at the problems of low accuracy and slow speed of road disease image detection,an improved SSD road disease detection algorithm is proposed.The algorithm uses the improved perception module to replace Conv4_3,FC7,Conv8,Conv9 and Conv10 layers in SSD network.Aiming at the problem of low recognition rate of pavement crack and depression image detection,hole convolution is used to expand the receptive field without losing resolution,so as to improve the feature extraction ability of shallow feature map for target objects.In order to make the network adaptively adjust the size of the receiving domain according to multiple scales of the input information,SKNet is introduced into the feature extraction layer.The experimental results show that the average accuracy of the improved method is 89.52% on the self-made road feature data set,and the detection speed is 56.23 FPS,which is 6.07%and 13.21 FPS higher than that of the original algorithm.The improved algorithm based on SSD can not only effectively identify the pavement damage features,but also effectively improve the recognition speed and efficiency.

关 键 词：路面破损 SSD算法 目标检测 Inception结构 EvoNorm SKNet 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]