Pre-sorting ore from waste material at the excavation face plays a key role in improving operational efficiency in open-pit mining operations. Automated solutions for this purpose need to be robust, operate in real time, and function on hardware with limited resources. This paper introduces a practical and reproducible framework for creating a lightweight ore classification system that can be deployed at the edge. The approach draws on a pre-trained YOLOv5s architecture, adapted via transfer learning to suit the industrial task using a public rock classification dataset that has been reformulated strategically. In fact, the model emerged from an accessible workflow that emphasizes efficiency. It underwent rigorous evaluation on an unseen test set, where it attained an overall accuracy of 80.7% and an AUC of 0.895. Notably, the model showed a recall of 88.4% for the economically important Ore class, which supports the main goal in industry of reducing product loss. Analysis of the results indicated that the model remains well calibrated. Classification errors tended to cluster around a handful of rock types that appear visually similar, which points to straightforward ways to refine it further. The main value of this work lies not in a new architecture but in a full end-to-end blueprint. This blueprint demonstrates the practicality of tailoring advanced object detectors to actual mining needs. Evidence from the study suggests that an optimized YOLO model offers a reliable and efficient option for automated quality control right at the mining face.
Zairov et al. (Mon,) studied this question.