To address the challenges of low efficiency and high computational costs faced by deep learning models in tunnel surrounding rock integrity analysis, a dual-model strategy based on YOLO and DeepLab is proposed, enabling rapid assessment of tunnel surrounding rock integrity. This method constructs a rock fracture recognition model based on the YOLO algorithm and introduces the DeepLab image semantic segmentation algorithm to intelligently extract fracture parameters, with results visualized. By integrating the processed image information, a new indicator for evaluating the development degree of rock fractures—the Fracture Factor (F_f)—is introduced to quantitatively analyze the development degree of surrounding rock fractures and determine the surrounding rock integrity coefficient (K_v). Using 5 000 images from typical domestic engineering projects for testing, the results show that the dual-model strategy based on YOLO and DeepLab achieves a fracture recognition accuracy of 96.41% and a fracture segmentation accuracy of 94.48%. Compared to traditional deep learning algorithms, the Fracture Factor (F_f) provides an assessment of surrounding rock integrity that is closer to the integrity coefficient in the BQ method. The dual-model strategy significantly reduces computational costs while ensuring accuracy, enhancing the feasibility of deep learning models in practical engineering applications.

By studying tunnel face images collected during a highway tunnel excavation, we establish the classification system of photolithography by MATLAB programming. According to the rock mass index (rock block size index or RBI) concept, 19 virtual lines are arranged in turn to the tunnel contour line, and the difference of the brazier rock mass structure is evaluated synthetically by 19 RBs of the radial direction of the face. Based on the structure, the distribution of comprehensive index of the rock mass structure (Z-RBI) is obtained through a comprehensive evaluation of the rock surface, and the corresponding relationship between the Z-RBI and rock structure type is determined. Considering the hardening degree of rock mass, the Z-RBI, groundwater condition, and initial geostress state, we propose the BP classification method for tunnel rocks, and we evaluated the BT method by using the analytic hierarchy process standard in an actual project. The comparison of the BT classification method and traditional BQ method reveals that the result of the former is more in accordance with the actual level of the excavated rock face than that of the latter method.

In order to realize the automatic recognition and classification of the surrounding rock lithology of the tunnel, a method of lithology recognition based on migration learning technology is proposed. First, pre-training on the Image-Net dataset by using the Inception-ResNet-V2 (IRV2) convolutional neural network model, and using model transfer learning technology to retrain the rock image dataset (including granite, limestone, basalt and shale) to obtain The lithology recognition model of the surrounding rock of the tunnel; then, the IRV2 model is tested, and the recognition performance of the three models: ResNet-50, Inception-V3 and VGG16 is compared; finally, the sub-image method and the overall image method are performed Comparative test of recognition effect. The experimental results show that: (1) The various classification performance indicators of IRV2 are all the best, and all can reach more than 90%, indicating that the model can realize the effective identification and accurate classification of surrounding rock lithology; (2) For rock pictures with more prominent texture, structure and structure, the recognition performance of the model is better; (3) The sub-image method can effectively improve the model’s performance compared to the overall image method. Identify performance.