为研究综合高效的隧道掌子面岩性智能分类方法，首先，通过收集高铁沿线施工隧道高清掌子面照片、地质素描图及工程地质说明，筛选并统计出灰岩、泥岩、砂岩、玄武岩4种岩性，在此基础上，采用图像增强扩充样本数量并构建岩性样本集； 然后，基于上述样本集分别构建ResNet50V2岩性分类迁移模型及VIT岩性分类模型，对比二者岩性分类效果，并采用Stacking方法集成2种模型的分类特点； 最后，通过对比3种元学习器（逻辑回归、支持向量机、决策树）对2种模型的集成融合效果来选取最适用的元学习器。结果表明：
采用逻辑回归集成ResNet50V2及VIT所构建的集成模型对岩性的分类效果最好，能充分融合掌子面岩性的全、局部特征来进行分类，模型准确率达到93.8%。
向露露
,
童建军
,
王明年
,
苗兴旺
,
叶沛
. Ensemble Convolutional Neural Networks and Visual Transformers: Research on Tunnel Face Rock Identification（集成卷积神经网络和视觉Transformer的隧道掌子面岩性判识研究）[J]. 隧道建设, 2024
, 44(5)
: 1056
-1067
.
DOI: 10.3973/j.issn.2096-4498.2024.05.013
The application of an increasing number of new
technologies in tunnel construction is a consequence of the rapid development
of intelligent tunnel construction. Among these, the use of machine vision
instead of human eyes for geological information recognition in tunnels has
become increasingly widespread. The authors build on this trend by using
machine vision to investigate intelligent classification methods for the
lithology of the tunnel face. First, four types of lithology (limestone,
mudstone, sandstone, and basalt) are screened and counted by collecting highdefinition
tunnel face photos, geological sketches, and engineering geological
descriptions along the highspeed railway. On this
basis, image enhancement techniques are used to expand the number of samples
and construct a lithology sample set. Then, based on the above sample sets, the
ResNet50V2 lithology classification transfer learning model and the VIT
lithology classification model are constructed. The lithology classification
effects of the two models are compared, and the Stacking method is used to
integrate the classification characteristics of the two models. The authors
select the most suitable metalearner by comparing the
integrated fusion effects of three metalearners,
logistic regression(LR), support vector machine, and decision tree, on the two
models. The results reveal that the Stacking model constructed by LR integrated
with ResNet50V2 and VIT has the best classification effect on lithology. Apart
from having an accuracy rate of 93.8%, it can also fully integrate the overall
and local features of the lithology of the tunnel face for better
classification.
