Improved mask R-CNN-based instance segmentation model for coal gangue.
Saved in:
| Title: | Improved mask R-CNN-based instance segmentation model for coal gangue. |
|---|---|
| Authors: | Sun, Peibo1 (AUTHOR) 1592984862@qq.com, Wu, Weimin1 (AUTHOR) |
| Source: | International Journal of Coal Preparation & Utilization. 2026, Vol. 46 Issue 6, p1601-1619. 19p. |
| Subject Terms: | *Image segmentation, *Coal mine waste, *Feature extraction |
| Abstract: | To achieve rapid sorting of coal gangue in mining environments, we propose an efficient model, RepSortNet, based on an improved Mask R-CNN to enhance the accuracy and reliability of coal gangue instance segmentation. First, to address the insufficient feature extraction caused by low contrast between coal and gangue, we construct an advanced backbone, RepViT-SG, which enhances the extraction of deep semantic features while maintaining real-time performance. Second, to improve the model's adaptability to gangue at various scales, we design the BiFPN-EMA module, which integrates multi-scale feature fusion with multi-level attention mechanisms, enhancing the model's ability to perceive critical information across channels. Finally, by incorporating the dynamic weight update strategy of the DyHead module, the model effectively enhances its generalization performance for detecting multi-appearance gangue in complex mining environments characterized by dynamic lighting, occlusion, and morphologically diverse targets. Experimental results demonstrate that the proposed model achieves an accuracy of 81.7% bbox_AP and 76.1% segm_AP on a self-constructed coal gangue dataset with a processing speed of 58 FPS, achieving a trade-off between detection speed and accuracy. [ABSTRACT FROM AUTHOR] |
| Database: | Energy & Power Source |
|
Full text is not displayed to guests.
Login for full access.
|
|
| Abstract: | To achieve rapid sorting of coal gangue in mining environments, we propose an efficient model, RepSortNet, based on an improved Mask R-CNN to enhance the accuracy and reliability of coal gangue instance segmentation. First, to address the insufficient feature extraction caused by low contrast between coal and gangue, we construct an advanced backbone, RepViT-SG, which enhances the extraction of deep semantic features while maintaining real-time performance. Second, to improve the model's adaptability to gangue at various scales, we design the BiFPN-EMA module, which integrates multi-scale feature fusion with multi-level attention mechanisms, enhancing the model's ability to perceive critical information across channels. Finally, by incorporating the dynamic weight update strategy of the DyHead module, the model effectively enhances its generalization performance for detecting multi-appearance gangue in complex mining environments characterized by dynamic lighting, occlusion, and morphologically diverse targets. Experimental results demonstrate that the proposed model achieves an accuracy of 81.7% bbox_AP and 76.1% segm_AP on a self-constructed coal gangue dataset with a processing speed of 58 FPS, achieving a trade-off between detection speed and accuracy. [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 19392699 |
| DOI: | 10.1080/19392699.2025.2505448 |
