基于改进U-Net的沥青拌合站混合料装车语义分割  

Semantic segmentation of mix loading at asphalt mixing plant based on improved U-Net

作　　者：李东丽 成高立 郭涛 夏晓华 Li Dongli;Cheng Gaoli;Guo Tao;Xia Xiaohua(Shaanxi Expressway Mechanization Engineering Limited Company,Xi'an 710038,China;Key Laboratory of Road Construction Technology and Equipment of MOE,Chang'an University,Xi'an 710064,China)

机构地区：[1]陕西高速机械化工程有限公司,陕西西安710038 [2]长安大学道路施工技术与装备教育部重点实验室,陕西西安710064

出　　处：《电子技术应用》2025年第4期29-34,共6页Application of Electronic Technique

基　　金：陕西省交通运输厅科研项目(24-74K);陕西高速机械化工程有限公司科技项目(KY01-24)。

摘　　要：针对现有的沥青拌合站混合料装车语义分割方法平均交并比(Mean Intersection over Union,mIoU)值较低、检测速度较慢等问题,提出一种轻量化网络RCS-UNet对沥青拌合站混合料装车状态进行语义分割。首先在U-Net网络中加入残差连接以缓解梯度消失的问题,使网络在训练过程中更加稳定,提高模型的收敛速度和泛化能力;其次加入坐标注意力(Coordinate Attention,CA)机制,增强位置与通道的信息感知,提高模型的特征提取能力,使模型更加关注图像中的重要区域;最后将U-Net网络中的标准卷积修改为深度可分离卷积,以减小模型的体积和参数量,使得模型在保持较高性能的同时,具有更低的资源消耗和更快的推理速度。实验结果表明,改进模型的准确率、mIoU值以及FPS值分别为99.20%、98.41%和22.98,与经典模型和当前先进模型相比三个指标均为最高,取得了最优的语义分割效果。Aiming at the existing asphalt mixing plant mixture loading semantic segmentation methods with low Mean Intersection over Union(mIoU)values and slow detection speed,a lightweight network RCS-UNet is proposed for semantic segmentation of asphalt mixing plant mixture loading state.Firstly,residual connections are integrated into the U-Net network to mitigate the gradient vanishing issue,promoting stability during training,enhancing convergence speed,and improving generalization abilities.Secondly,the Coordinate Attention(CA)mechanism is incorporated to boost the perception of positional and channel information,refining feature extraction and enabling a sharper focus on critical regions within the image.Finally,the standard convolution in the U-Net network is modified to depth-separable convolution in order to reduce the size and parameters of the model,so that the model has a lower resource consumption and a faster inference speed while maintaining a higher performance.The experimental results show that the accuracy,mIoU,and FPS of the improved model are 99.20%,98.41%and 22.98,respectively,which are the highest compared with the classical model and the current state-of-the-art model.The best segmentation results are obtained.

关 键 词：残差连接 CA 深度可分离卷积 语义分割 沥青拌合站 

分 类 号：U415[交通运输工程—道路与铁道工程]