Abstract: [Objective] To overcome the limitations of conventional manual methods for rock fracture identification—such as low efficiency, high subjectivity, and limited accessibility in rugged terrain—this study aims to develop an approach for rapid and accurate fracture recognition and parameter extraction, particularly on steep rock slopes.[Methods]An enhanced U-Net model was developed and trained on the publicly available GeoCrack dataset. To better capture the irregular, linear characteristics of fractures, the model integrates a Convolutional Block Attention Module (CBAM) and a multi-scale feature fusion mechanism. The AdamW optimizer combined with a cosine annealing learning rate scheduler was employed to accelerate convergence and ensure training stability. Recognized fractures were refined using post-processing techniques, including Gaussian blur, morphological operations, and skeletonization. Fracture characteristic parameters were then calculated by integrating the 2D image data with 3D point clouds using camera parameters. The proposed workflow was validated using drone-captured imagery of the Jigongyan rock mass. [Results]Experimental results demonstrate that the improved U-Net model outperforms both a traditional Fully Convolutional Network (FCN) and the original U-Net in terms of Pixel Accuracy (PA), Mean Pixel Accuracy (MPA), and Mean Intersection over Union (MIoU). In the Jigongyan case study, the dominant fracture orientations identified by the model (approximately 320° and 140°, with dip angles of 75°–85°) show strong agreement with field mapping data (e.g., T1: 330°∠82°; T3: 170°∠82°). The calculated 3D fracture length and width exhibited minimal errors. [Conclusion]This study presents an automated workflow for rock fracture identification and parameter quantification. The method not only reduces survey costs and improves accuracy but also provides a reliable reference for designing engineering mitigation measures, demonstrating considerable practical value.