High-Precision Detection of Core Fractures Based on Multi-Scale Feature Fusion and Spatial Compression

High Resolution Image Download MS PowerPoint Slide Natural fractures play a critical role in controlling the pore structure, fluid flow capacity, and reservoir heterogeneity in low-permeability oil and gas reservoirs. However, their subtle morphology, wide scale variation, and complex background conditions render traditional manual identification methods inefficient and inherently subjective. To address these challenges, this study proposes an improved model, YOLOv11-EMBSFPN-SC-AP (YOLOv11-ESA), and develops a lightweight, high-precision intelligent framework for rock-fracture detection. In the backbone network, the C3k2 Asymmetric Padding (AP) module is incorporated to enhance local and global feature extraction, while the C2PSA (position-sensitive attention) mechanism enables two-stage attention fusion that significantly improves the representation of fine fracture details. The SPPF (Spatial Pyramid Pooling Fast) module is further employed to strengthen the multiscale robustness of deep features. In the neck, an Enhanced Multi-Branch Spatial Feature Pyramid Network (EMBSFPN) is constructed by combining heterogeneous convolution kernel selection, weighted multiscale fusion, and the Efficient Upsampling Convolution Block (EUCB) to achieve efficient feature transmission under lightweight constraints. A Spatial Compression (SC) module is additionally integrated to remove redundant information and highlight key fracture regions. In the detection head, the Asymmetric Padding (AP) module is introduced to enhance multiscale prediction layers, ensuring high accuracy across fractures of varying sizes. Experiments conducted on a data set of 4209 core images demonstrate that YOLOv11-ESA achieves an mAP00.5 of 82.6%, an mAP00.5:0.95 of 65.1%, and a recall of 75.0%, representing improvements of 0.4, 0.8, and 0.4% over the baseline YOLOvl1, respectively. Notably, the model attains an inference speed of 1096.5 Frames Per Second (FPS), which is 98.6% faster than the original YOLOvl1, while maintaining a compact model size of 4.38 MB with only 2.05 M parameters. These results indicate that the proposed model achieves an optimal balance among detection accuracy, real-time performance, and a lightweight design. This study provides a robust technical foundation for fine reservoir characterization and intelligent core analysis and can be broadly applied to automated fracture detection in oil and gas exploration and laboratory scenarios.