Multi-Level Coal Density Classification via Cross-Modal Interaction With Structure–Texture Feature Fusion
IEEE Access · 2026 · 14 · 59279-59293
DOI: 10.1109/access.2026.3678527 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Shuxian Su und 4 weiteren Autor:innen. Die Arbeit nennt 32 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: Natural Science Foundation of Jiangsu Province.
Abstract (Originalsprache)
With the growing demand for efficient coal utilization, the intelligent multi-level classification of coal density has emerged as a significant research focus. Existing approaches primarily rely on optical (multispectral, infrared, visible-light) or X-ray data with deep learning; however, optical modalities are sensitive to environmental conditions and capture only surface cues, while X-ray imagery reflects average internal properties but lacks spatial density distributions. Consequently, prior work largely remains at binary coal–gangue discrimination, limiting fine-grained tri-class separation of clean coal, middlings, and gangue. We propose a Cross-Modal Interaction Enhancement Model (CMIEM) that fuses surface morphology from visible-light images with internal structure from CT, overcoming the information constraints of single modalities. Addressing the limitations of traditional multimodal fusion (e.g., feature concatenation or weighted averaging) and the weak transfer of cross-domain models (e.g., XFMamba), CMIEM redesigns the fusion stage with a Multi-Head Cross-Attention (MHCA) mechanism to capture fine-grained bidirectional dependencies, and a Cross-Modal Enhancement Module (CMEM) that integrates layer normalization, feed-forward networks, and adaptive gating for hierarchical, dynamic fusion. On a self-built tri-class dataset (clean coal–middlings–gangue), CMIEM attains 94.15% accuracy, outperforming the strongest baseline by 9.24 percentage points. Ablation studies show complementary gains: using only MHCA or only CMEM yields 90.91% accuracy, whereas the full model reaches 95.45%. Compared with advanced multimodal classifiers (Effective-MMDF, Local-to-Global, XFMamba), CMIEM achieves the best performance across density ranges, with pronounced improvements on the middlings class. These results demonstrate the value of CT–visible fusion for fine coal separation and provide a practical pathway toward robust recognition of composite particles in mineral processing.
Förderung
- Natural Science Foundation of Jiangsu Province
Zitieren
APA
Su, S., Geng, P., Zhang, G., Xue, G., & Dong, L. (2026). Multi-Level Coal Density Classification via Cross-Modal Interaction With Structure–Texture Feature Fusion. IEEE Access, 14, 59279-59293. https://doi.org/10.1109/access.2026.3678527
Vancouver
Su S, Geng P, Zhang G, Xue G, Dong L. Multi-Level Coal Density Classification via Cross-Modal Interaction With Structure–Texture Feature Fusion. IEEE Access. 2026;14:59279-59293. doi:10.1109/access.2026.3678527
BibTeX
@article{Su2026,
author = {Su, Shuxian and Geng, Peixian and Zhang, Gansu and Xue, Gehang and Dong, Liang},
title = {Multi-Level Coal Density Classification via Cross-Modal Interaction With Structure–Texture Feature Fusion},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {59279--59293},
doi = {10.1109/access.2026.3678527}
}
EUVIMED