摘要
为提高自密实混凝土骨料语义分割算法的性能,本文研究了一种融合注意力机制与深度学习的计算机视觉方法,该方法采用主流的编码器(ResNet50)-解码器(U-net)结构,建立了基于融合注意力机制与深度学习的自密实混凝土骨料语义分割模型,可以在像素级别上分割混凝土图像中的骨料,并通过精度、召回率、F1分数和交并比等4个指标对模型进行评估。实验结果表明,在同一数据集下,本文模型评估性能优于目前性能优异的模型DeepLab V3+、PSPNet和HRnet,为评价自密实混凝土的稳定性或抗静态离析性提供了高效率的工具。
To improve the performance of semantic segmentation algorithm for self-compacting concrete aggregate,a computer vision method integrating attention mechanism and deep learning was developed.This method adopted the mainstream encoder(ResNet50)-decoder(U-net)structure,and established a semantic segmentation model of self-compacting concrete aggregate based on attention mechanism and deep learning.It could segment aggregate in concrete image at pixel level.The model was evaluated by four indicators:precision,recall,F1 score and IoU.The experimental results show that under the same dataset,the evaluation performance of this model is better than those of DeepLab V3+,PSPnet and HRnet,which have excellent performance at present.This provides an efficient tool for evaluating the stability or static segregation resistance of self-compacting concrete.
作者
崔李三
邓鹏
周圆兀
CUI Lisan;DENG Peng;ZHOU Yuanwu(School of Civil Engineering and Architecture,Guangxi University of Science and Technology,Liuzhou 545006,China)
出处
《广西科技大学学报》
CAS
2023年第4期40-47,共8页
Journal of Guangxi University of Science and Technology
基金
广西高校中青年教师科研基础能力提升项目(2022KY0348)
国家自然科学基金项目(51908141)
广西科技大学博士挂职驻柳企业工作项目(BSGZ2127)
广西科技大学博士基金项目(校科博14z13)
广西高等教育本科教学改革工程项目(2020JGZ129)
2022年度校级本科教育教学改革项目(2022XJJG51)资助。
关键词
深度学习
注意力机制
语义分割
自密实混凝土
稳定性评价
deep learning
attention mechanism
semantic segmentation
self-compacting concrete
stability evaluation
