Existing algorithms for recognizing rock discontinuities exhibit several limitations, including inadequate universality across diverse point cloud features, low computational efficiency, and limited applicability in engineering practice. To address these issues, an intelligent method based on neural networks is proposed. First, raw point cloud data are standardized through preprocessing, and representative feature regions are manually selected to construct a high-quality training dataset. Second, the clustering by fast search and find of density peaks (CFSFDP) algorithm is employed to generate sample labels. Third, a multilayer perceptron (MLP) model and a multilayer convolutional neural network (MCNN) model are constructed and trained. These models use point normals from the full-scale point cloud as input for preliminary discontinuity classification, and their performance is comparatively analyzed. Finally, the hierarchical density-based spatial clustering of applications with noise (HDBSCAN) algorithm is applied to refine the classification results and to calculate the spatial orientations of the discontinuities. Experimental and comparative results demonstrate that (1) the MLP model provides higher processing speed for simple discontinuity structures, whereas the MCNN model achieves higher classification accuracy for complex and nonuniform point clouds; and (2) compared with traditional clustering methods, the proposed approach reduces computational time by 25%－50% while effectively overcoming the limitations of conventional algorithms in heterogeneous geological environments, demonstrating strong robustness.