Abstract: Semantic segmentation methods exhibit limited accuracy in the individual segmentation of rock chips due to dense stacking, large size variations, and severe edge occlusion in rock chip images collected during shield tunneling. To address these challenges, a rock chip geometric feature recognition model based on an improved mask region-based convolutional neural network (Mask R-CNN) is proposed. First, a rock chip image dataset is constructed, including laboratory (dispersed and stacked) and field scenarios from different strata of a tunneling project in Beijing, and data augmentation is applied to expand the sample size. Through comparative experiments, Mask R-CNN is selected as the baseline model. An efficient multiscale attention (EMA) mechanism is introduced to enhance the backbone network′s ability to extract multiscale features. Additionally, the Confluence-non-maximum suppression (Confluence-NMS) algorithm replaces the traditional non-maximum suppression method to reduce false detections and missed detections in dense scenarios. Ablation experiments conducted on the engineering dataset show that (1) after introducing EMA and Confluence-NMS, the model achieves AP_0.50, AP_0.75, and AP_0.50∶0.95 values of 82.70%, 73.69%, and 59.13%, respectively, representing improvements of 4.93%, 3.19%, and 2.60% over the baseline model, while the F₁-score increases by 3.63%; (2) significance tests confirm the statistical effectiveness of the proposed improvements. Based on the segmentation masks, the rotating calipers algorithm, combined with a reference-object conversion method, is used to automatically extract geometric features of rock chips, including the major and minor axis lengths, area, and perimeter. On the engineering test set, the particle size distribution predicted by the model is highly consistent with that obtained from manual annotations: the major axis lengths are primarily concentrated in the 25－75 mm range, while the minor axis lengths are mainly distributed within the 15－45 mm range. The maximum error in the proportion of each interval is less than 7.53%, with slightly higher errors occurring only in the small-size interval (<25 mm) due to missed detections caused by dense occlusion. In the evaluation of generalization performance, the model achieves an AP_0.50∶0.95 of 57.62% on rock chip images from a tunnel project in Qingdao, demonstrating its cross-project applicability.

Key words: shield tunnel, rock chip morphology distribution, deep learning, instance segmentation, mask region-based convolutional neural network