OGI-RT-DETR: A Lightweight Vision-Sensing Model for Real-Time Graphite Ore Grade Recognition in Intelligent Sorting

Area of Science:

• Computer Vision
• Artificial Intelligence
• Materials Science

Background:

• Vision-based sensing is crucial for real-time ore grade recognition in intelligent mineral sorting.
• Industrial graphite ore images suffer from illumination variation, background interference, and subtle grade differences, hindering model reliability on edge devices.

Purpose of the Study:

• To propose OGI-RT-DETR, a lightweight, real-time, end-to-end vision-sensing model for accurate graphite ore grade recognition.
• To enhance feature extraction, multi-scale semantic interaction, and localization stability for robust performance in industrial settings.

Main Methods:

• Developed OGI-RT-DETR based on RT-DETR-r18, incorporating a PConv-Rep backbone for efficient feature extraction.
• Implemented a CFPT feature fusion module to improve multi-scale semantic interaction and spatial fusion.
• Introduced Wise-Focaler-MPDIoU loss for enhanced bounding box regression accuracy and stability.

Main Results:

• OGI-RT-DETR achieved 85.6% Precision, 87.6% Recall, 86.1% mAP50, and 84.5 FPS on a self-built dataset.
• Outperformed baseline RT-DETR-r18 by improving mAP50 by 3.7% and FPS by 8.2%.
• Reduced computational load with 31.8% fewer FLOPs and 38.2% fewer parameters.

Conclusions:

• OGI-RT-DETR offers an accurate, efficient, and lightweight solution for real-time graphite ore grade recognition.
• The model demonstrates superior performance and reduced resource requirements, suitable for edge device deployment in intelligent mineral sorting.