<p>Accurate characterization of geological conditions at the tunnel face is fundamental for determining the optimal excavation techniques in tunnel engineering. The joints on a tunnel face can be used to characterize the geological strength of the rock surrounding the tunnel. Traditional geological strength index (GSI) is assessed by engineers based on their experience through observing the tunnel face, but it lacks objectivity and accuracy. Therefore, a fully automated algorithm is needed to improve the detection efficiency. In this study, we used a hybrid of the U-Net and multilayer perceptron (MLP) models to automatically predict the GSI from the tunnel face rock joint images. First, the U-Net model was improved using depthwise separable convolution and a dual attention module, which can effectively detect the semantic information of the fractures in the tunnel face. Then, the geometric parameters of the rock joints were calculated based on the semantic segmentation images, and the MLP model was used to predict the GSI of the tunnel face. Finally, taking the accuracy, recall rate, F₁ score, and mean intersection over union score as evaluation indicators, the proposed algorithm was compared with existing algorithms. A mobile application was designed to import the trained model into a mobile device for verification. The experimental results revealed that the detection accuracy of the newly developed method was 92%, and the detection time was 0.0075&#xa0;s. Compared with the traditional methods, the amount of data was reduced by 68.9%, and the amount of computation was reduced by 80.2%. Moreover, the accuracies of the recognition of joints in the tunnel face and the GSI prediction were improved. The proposed method provides a basis for roadway support and excavation. By automatically obtaining rock parameters by taking pictures with the mobile application, the work efficiency of geological engineering personnel can be enhanced, and greater convenience and timeliness can be achieved.</p>