基于卷积神经网络的碎石撒布状态检测算法优化研究  

Researchon Optimization of Chip Spreading State Detection Algorithm Based on Convolutional Neural Network

作　　者：陈嘉毅 韩磊[1] 叶敏[1] 左献宝 郝长峰 

机构地区：[1]长安大学公路养护装备国家工程实验室,陕西西安710064 [2]河南省高远公路养护技术有限公司 [3]河南省高等级公路检测与养护技术重点实验室

出　　处：《工程机械》2025年第5期62-66,I0004,共6页Construction Machinery and Equipment

基　　金：河南省重点研发专项(231111520200);河南省重点研发专项(241111241600);新乡市重大科技专项(22ZD013)。

摘　　要：针对卷积神经网络在同步碎石封层碎石撒布状态检测中面临检测精度不足与计算复杂度过高的问题,提出一种基于深度卷积的卷积神经网络优化策略。首先利用拍摄的同步碎石封层施工现场图像,制作碎石撒布状态数据集。随后,使用皮尔逊相关系数分析碎石撒布图像,发现碎石撒布图像各通道间存在较高的线性相关性,并针对此现象提出卷积神经网络优化策略。最后,采用优化策略优化ResNet50、ResNet101和ResNet152,并使用全卷积网络作为语义分割头部,与未优化前的模型进行性能对比。同时,注册前向钩子函数提取第一次残差连接后生成的全部特征图,使用皮尔逊热力图分析任意两张特征图之间的线性关系。结果表明:所提出的卷积神经网络优化策略能够降低模型训练过程中的通道特征穴余。经过优化后的ResNet50、ResNet101和ResNet152模型,计算复杂度分别降低40.3%、45.8%和47.9%,平均交并比分别提升0.23%、0.23%和0.34%,验证了所提出的优化策略在提升模型性能方面的有效性。In view of the problems of insufficient detection accuracy and high computational complexity faced by the convolutional neural network in the detection of chip spreading state of synchronous chip sealing,a convolutional neural network optimization strategy based on deep convolution is proposed.Firstly,the captured images of synchronous chip sealing construction site are used to create a chip spreading state dataset.Subsequently,the chip spreading image is analyzed using Pearson correlation coef-ficient,it is found that there is a high linear correlation among the channels of the chip spreading image,and a convolutional neural network optimization strategy is proposed for this phenomenon.Finally,the optimiza-tion strategy is used to optimize ResNet50,ResNet101 and ResNet152,the full con-volutional network is used as the semantic segmentation head,and the performance is compared with that of the models before optimization.Meanwhile,the forward hook function is registered to extract all the fea-ture maps generated after the first residual connec ction,and the linear relationship be-tween any two feature maps is analyzed us-ing Pearson heatmap.The results show that the proposed convolutional neural network optimization strategy can reduce the chan-nel feature redundancy during model train-ing.The optimized ResNet50,ResNet101 and ResNet152 models have a reduction in computational complexity by 40.3%,45.8%and 47.9%,respectively,and an increase of mean intersection over union by 0.23%,0.23%and 0.34%,respectively,which verifies the effectiveness of the proposed optimization strategy in terms of improving model performance.

关 键 词：碎石撒布状态检测 卷积神经网络 语义分割 数字图像处理技术 

分 类 号：U416.2[交通运输工程—道路与铁道工程]