Stellar-YOLO: A Graphite Ore Grade Detection Method Based on Improved YOLO11

Mineral recognition technology is crucial for improving mining efficiency and advancing smart mining development. To enable the efficient deployment of graphite ore grade detection on edge computing devices, we propose Stellar-YOLO, a YOLO11-based detection framework with asymmetrical architecture optimizations tailored for real-world conditions. The backbone is replaced by the lightweight StarNet to enhance computational efficiency, while the C3k2-CAS module, integrating convolution and additive attention, is embedded in the neck to improve feature expressiveness. The head incorporates the SEAM module, forming the Detect-SEAM, to boost the recognition of complex mineral details. Moreover, to robustly adapt to real mining environments, we apply simulated data augmentation techniques involving motion blur, dust noise, and low brightness conditions. Stellar-YOLO achieves 93.6% mAP based on a custom-built graphite ore dataset, outperforming the baseline by 4.5% and reducing the FLOPs, parameters, and model size by 27%, 26%, and 23%, respectively. This work explores how asymmetrical architectural innovations and robustness-oriented evaluation contribute to a lightweight and effective approach for computer vision-based mineral quality assessment, demonstrating strong potential for practical applications in real-world industrial environments.