A Lightweight Rock Core Image Instance Segmentation Method Based on MCE-YOLO11n-seg

Abstract In hydraulic engineering, manual rock core logging and rock quality designation (RQD) calculation exist the problems of substantial workloads, low efficiency, and high subjectivity. To address these issues, a lightweight instance segmentation model, termed MCE-YOLO11n-seg, was proposed for the automatic logging of rock cores in core box images and automated RQD calculation. Base on YOLO11n-seg, MobileNetV4 was adopted to replace the original backbone, thereby reducing computational complexity and the number of parameters. Additionally, a content-aware reassembly of features (CARAFE) module was introduced to optimize multi-scale feature fusion. Furthermore, an efficient enhanced segmentation head (EES-Head) was developed to sharpen the representation of rock core instance boundaries by utilizing an edge branch and an edge-aware feature modulation mechanism. The results show that MobileNetV4 enhances the lightweight of the model, and CARAFE and EES-Head effectively improve multi-scale feature fusion and boundary segmentation. MCE-YOLO11n-seg achieved 1.89 M parameters, 5.88 G FLOPs, and an inference speed of 89.81 FPS. The proposed model yielded a Precision of 98.18%, a Recall of 98.00%, and mask [email protected]:0.95 of 88.03%. Compared with YOLO11n-seg, it reduced parameters and FLOPs by 33.47% and 38.61%, respectively, while increasing the mask [email protected]:0.95 by 0.52 percentage points. Practical validation across 23 core box images demonstrated a mean relative error of merely 1.71% and a median relative error of 1.68% between the automated RQD calculations and manual measurements. MCE-YOLO11n-seg achieves a balance between lightweight design and high segmentation accuracy, offering effective support for automated RQD calculation and intelligent core logging.