Fig. 1. Deformable convolution in ResNet-101. (a) Architecture of ResNet-101. (b) Comparison of the two convolutional kernels in fracture feature extraction: b-1 standard convolution; b-2 deformable convolution. (c) Illustration of 3 × 3 deformable convolution.
Fig. 2. Architecture of the YOLACT++ model in fracture detection.
Fig. 3. Definition of the IoU.
Fig. 4. Fracture quantification. (a) Image merging: a-1 and a-2 fracture local image; a-3 fracture panorama. (b) Extraction of the fracture skeleton. (c) 3D point cloud of the fracture.
Fig. 5. Dataset acquisition and preprocessing. (a) Dataset acquisition: a-1 overview of the dataset acquisition; a-2 proximity photogrammetry. (b) Color adjustment: b-1 original; b-2 brightness modifying; b-3 grayscale modifying. (c) Multi-angle rotations: c-1 36°; c-2 72°; c-3108°; c-4144°. (d) Mirroring: d-1 horizontal mirroring; d-2 vertical mirroring; d-3 central symmetry.
Fig. 6. Loss curves of the YOLACT++ model.
Fig. 7. The performance evaluation metric curves during the training process of YOLACT++, Mask R-CNN and YOLO V8. (a) IoU. (b) Precision. (c) Recall.
Fig. 8. Scatter box plots showing the performance of YOLACT++, Mask R-CNN and YOLO V8. (a) ResNet-50. (b) ResNet-101.
Fig. 9. Segmentation results of YOLACT++, Mask R-CNN and YOLO V8 for fractures in different scenarios.
Fig. 10. Applications of Achievements in Engineering Geology. (a) Rock fractures with complex shapes and branching. (b) The YOLACT++ identification result. (c) Fine mesh model for rock fractures. (d) Numerical simulation of the SIF and fracture propagation path.