岩石岩性轻量化网络识别方法及其隧道工程应用

Lightweight network identification method of rock lithology and its application in tunnel engineering

在隧道工程地质勘察中,需要对岩石岩性进行判断和识别,以便开展工程岩体定量分析及工程地质安全评价。鉴于目前岩石岩性识别过程中存在主观性强、周期性长等问题,结合岩石表面岩性特征信息,提出基于MobileNetV3轻量化网络模型的岩石岩性快速识别方法。首先采集工程中常见的20类不同岩性的岩石图像样本并搭建图像数据集,通过构建MobileNetV3网络模型和迁移学习方法对岩石图像进行训练,获得MobileNetV3岩石岩性识别模型;然后,将MobileNetV3岩性识别模型与EfficientNet-B0,Xception和Inception-ResNet-v2网络模型的岩性识别结果进行对比,从而评估其训练效果。此外,采用MobileNet网络替换SSD模型的主干特征提取网络,构建轻量化MobileNet-SSD目标检测模型对岩石图像进行检测与识别。结果表明,MobileNetV3网络模型识别岩石岩性的准确率为98.2%,模型大小仅为12 MB,对测试集上单张图像从输入到输出识别结果的平均时间为812 ms;与其他模型相比,该模型在识别精度、模型大小和识别时间上,都具有一定优势;通过MobileNet-SSD目标检测模型,可以实现对岩石目标的定位以及多种岩石同时识别。将MobileNetV3网络模型应用于隧道掌子面岩石岩性识别中,对不同岩性以及同一岩性不同完整程度的隧道掌子面岩石,模型有较好的识别效果。本研究成果有助于提高卷积神经网络模型的可移植性以及隧道工程离线环境下岩石岩性的快速准确识别。

In the tunnel engineering geological survey,it is necessary to judge and identify the rock lithology in order to carry out quantitative analysis of engineering rock mass and engineering geological safety evaluation.In view of the problems of strong subjectivity and long periodicity in the process of rock lithology identification,a fast identification method of rock lithology based on MobileNetV3 lightweight network model was proposed by combining the lithology characteristics of rock surface.Firstly,20 kinds of rock image samples with different lithology were collected and the image data set was built.The rock image was trained by constructing MobileNetV3 network model and transfer learning method to obtain MobileNetV3 rock lithology recognition model.Then,the MobileNetV3 lithology identification model was compared with the lithology identification results of the EfficientNet-B0,Xception,and Inception-ResNet-v2 network models to evaluate its training performance.In addition,the MobileNet network was used to replace the backbone feature extraction network of the SSD model,and a lightweight MobileNet-SSD target detection model was constructed to detect and identify rock images.The results show that the accuracy of identifying rock lithology by MobileNetV3 network model is 98.2%,the model size is 12 MB.The average time from input to output for a single image on the test set is 812 ms.Compared with other models,this model has certain advantages in recognition accuracy,model size and recognition time.The MobileNet-SSD target detection model enables the positioning of rock targets and simultaneous identification of multiple rocks.The MobileNetV3 network model is applied to the rock lithology identification of tunnel face.The model has a good identification effect for the rock of the tunnel face with different lithology and different integrity of the same lithology.The research results are helpful to improve the portability of convolutional neural network model and the fast and accurate identification of rock lithology in the offline environment of tunnel engineering.