Abstract: The characteristics of rock fractures are the core indices for evaluating the structure and integrity of rock masses, and are also the important factors for evaluating the safety and stability of rock engineering. For the rock fracture recognition based on deep learning, the Unet model is improved by introducing the convolutional block attention module, effectively improving the accuracy of rock fracture recognition. For separation and feature extraction of intersection fractures, a fracture separation and characterization algorithm based on the trace direction judgment is proposed. According to the characteristics of fracture separation, the overlapping tracing method or fracture tracing method is used to separate the intersection fracture skeleton. Then, the differential accumulation method, box method and linear regression method are employed to calculate the geometric characteristic indices such as the length, width and dip angle of the fractures. Based on the proposed algorithm, a set of software system for the intelligent identification and characterization of rock fracture images with a graphical user interface is developed, achieving complete functions from deep learning model parameter selection, model training, fracture recognition, quantitative analysis to result visualization. Finally, the performance of rock fracture recognition and separation characterization algorithms is evaluated. The results show that the improved Unet model has the best recognition performance for complex distribution fractures, and its overall recognition performance is superior to that of other networks. The skeleton separation algorithm achieves the expected results for the separation of common types of intersection fractures, and the characterization algorithm has high accuracy in characterizing the separation and intersection fractures. It has good application effects on actual rock fracture images. The research can provide reference for rock engineering tests and detection of rock mass structure by using the computer vision.