Advanced Fracture Segmentation Techniques for Enhancing Intelligent Detection of Coal Roadway Roof Strata

Abstract

The identification of rock fractures in strata is crucial to enhance the intelligence of rock detection. Traditional fracture feature extraction methods suffer from issues such as low accuracy and low processing speed, necessitating the development of more effective approaches. To address this problem, this study proposes a new fracture instance segmentation network called FracSeg. Based on the SOLOv2 framework, we incorporated the Swin Transformer to optimize the backbone network and enhance fracture feature extraction. The CARAFE operator is utilized to replace nearest neighbor interpolation, reducing the computational overhead when merging multi-scale fracture features. Finally, the Shuffle Attention module was employed to improve the network’s detection of fracture features. The experimental results demonstrate the superior performance of FracSeg, achieving a mask mAP of 78.2 on a custom dataset while maintaining an average inference speed of 28.2 fps. Even under complex conditions, it outperformed previous fracture segmentation networks in identifying crack structures in coal roadway roofs. Additionally, ablation studies verified the effectiveness of each optimized component in the FracSeg model.

Keywords

• Instance segmentation
• borehole image
• fracture recognition
• SOLOv2