To more efficiently determine the integrity degree of tunnel surrounding rock, a precise determination method for the integrity degree of tunnel surrounding rock based on lightweight neural network MobileNet-v2 is proposed. Firstly, grayscale the image, denoise the image, and detect the edges of cracks; Then, the MobileNet-v2 lightweight neural network model is pre trained on the ImageNet dataset, and combined with transfer learning to complete data detection on the training, validation, and testing sets; Finally, a comparative experiment was conducted with traditional neural networks RestNet-50 and VGG16. By identifying the area, width, and length of cracks, the crack ratio Ks is introduced as an indicator to evaluate the integrity of surrounding rock. The results show that: (1) In terms of accuracy, loss value, and training time, the MobileNet-v2 model is significantly better than the VGG16 and RestNet-50 models in this experiment. (2) The MobileNet-v2 model has the highest accuracy, with a validation set accuracy of around 94%. (3) By comparing with the results of on-site experiments, it has been proven that using digital image processing methods to evaluate the integrity of rock masses has high accuracy and feasibility. The research results provide a reference basis for more accurate determination of the degree of rock integrity.