Abstract:
In order to solve the problems of long identification time, low security, and high subjectivity in the existing identification methods of tunnel lithology, given the fact that composition characteristics differ among lithological surfaces, a rapid identification method of tunnel lithology based on the lightweight model and rock images was proposed. First, six types of major rocks in tunnels, including gneiss, granite, limestone, marble, tuff and sandstone, were collected by camera, and the rock image data set was established and divided into training set, verification set and test set. Then, based on the lightweight model MobileNet V2, pre-training was conducted on the ImageNet data set, the structure of the model classifier was improved to adapt to the rock data set, and 1170 images of the training set were trained using the transfer learning method for model training to obtain the rock lithology recognition model. Finally, a total of 300 test set images were selected and tested offline, and compared with those of the VGG16 model and the SVM (support vector machine) model. The experimental results show that the overall evaluation indexes of the model on the test data set were above 85%, of which the evaluation indexes of tuff reached more than 94%, the size of the model was only 28.3 MB, and the average recognition time was 2880 ms, indicating that the recognition model was small in size, high in recognition accuracy, and fast in recognition time, which is superior to traditional methods in accuracy and recognition speed.