Abstract: The manual interpretation of structural planes in rock masses during tunnel excavation and construction is highly subjective and inefficient. To address this, computer vision technology is applied to identify structural planes and analyze their geometric parameters. First, an improved neural network model is improved to segment the contours of tunnel faces and structural planes. Next, five base models are integrated using two ensemble learning methods. Finally, image processing techniques are used to analyze the geometric parameters of structural planes, including the number of structural planes, the number of groups, and the average distance. The results indicate the following: (1) The precision, recall, and f₁ scores of the U²-Net semantic segmentation model are 0.929，0.926，and 0.928. (2) Five improved base models, derived from the Swin Transformer model, are utilized for structural plane segmentation. The I_IoU and M_IoU of the SwinT-Upernet base model reach 66.95% and 81.72%, respectively, while the M_Acc of the Resnet50-Upernet base model reaches 88.33%. (3) The accuracy of structural plane group number analysis using the weighted average method is 82.9%, and the relative error in the average distance analysis is 16.15%.

Key words: tunnel engineering, structural plane geometric parameter, deep learning, rock mass structural plane, tunnel face image