Abstract: Rock Quality Designation (RQD) is a commonly used indicator for rock mass classification and evaluation in the geological industry. Manual acquisition of RQD is cumbersome and subjective. A framework for intelligent RQD recognition and computation was proposed, utilizing an improved Mask R-CNN two-stage instance segmentation deep learning network for automated RQD computation. Improvements to Mask R-CNN included adding clustering algorithms for prior clustering of the dataset, replacing the backbone network with ResNext, and adding path aggregation networks. Ablation experiments indicated the effectiveness and reliability of the improvements. The framework obtained good feature extraction capabilities through transfer learning on a small sample dataset. The intelligent recognition and computation tasks for RQD were divided into three subtasks: core box segmentation, single row core segmentation, and core segment segmentation, extracting core box, single row core, and core segment information from borehole core images, thereby achieving automated RQD computation. Automated RQD computations were performed on the boreholes of the Shituo Yangtze River Bridge in Shituo Town, Fuling District, Chongqing. The results showed that the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) were 3.497 and 4.654,respectively, indicating the framework's good accuracy, efficiency, and generalization ability. Factors affecting the framework's predictive results were discussed, and possible directions for further research were explored.