To address the issues of traditional tunnel surrounding rock classification, such as over-reliance on manual experience, strong subjectivity, and low efficiency, this study innovatively introduces the Vision Transformer model and constructs a full-process real-time surrounding rock identification system with the workflow of "acquisition-processing-evaluation". Breaking through the limitation of the local receptive field of traditional convolutional neural networks, the system achieves efficient capture of global geological features. It comprises three components: a digital image acquisition terminal, an intelligent classification processing module, and an intelligent evaluation module. Based on a dataset of 2,621 tunnel face images covering multiple geological conditions, the model reaches a recognition accuracy of 92.49% after 100 training epochs, outperforming the EfficientNet and ResNet34 models. Through a lightweight mobile App, rapid image acquisition, intelligent classification, and real-time recommendation of construction methods can be realized. Verified in the Xiaoyuan Tunnel project, the system shows a consistency of over 90% with the results of the traditional BQ classification method. The recommended construction methods and blasting parameters effectively reduce over-excavation and under-excavation deviations as well as surrounding rock disturbance, providing an innovative and reliable technical solution for intelligent tunnel construction.