Basalt visible light image recognition optimization algorithm based on YOLOv8

With the rapid advancement of photoelectric intelligent sorting technology, YOLOv8-based algorithms for scheelite recognition have been successfully optimized. However, the development of scheelite resources is often accompanied by substantial waste rock accumulation and tailings discharge, which pose significant environmental challenges. This study focuses on the waste rocks obtained after the pre-concentration stage of scheelite processing. Due to the complex surface features of basalt and its high visual similarity to non-target gangue in visible light images, accurate identification remains difficult. To address this issue, we propose the YOLOv8-Basalt algorithm, which integrates multiple optimization strategies including HSV-based image enhancement, Inner-SIoU loss function, ECA attention mechanism, TSiLU activation function, Optuna-based hyperparameter tuning, and GhostConv lightweight convolution.Ablation and comparative experiments were conducted to evaluate the independent contributions of each module to recognition accuracy and inference efficiency. As a result, the average precision for basalt identification increased from 0.922 to 0.979, and the F1-score improved from 0.90 to 0.93, while maintaining a low inference time of only 2.60 ms. This demonstrates the algorithm’s ability to achieve accurate and rapid recognition of basalt in visible light images.Considering the complexity of deploying YOLO-based models on AI accelerators and their limited compatibility with industrial color sorters, we introduce a simulation-based waste rejection prediction method as an alternative. This approach enables rapid evaluation of sorting performance without relying on specific hardware. Based on quantitative simulation experiments, the predicted waste rejection rate for the test set samples reached 66.46%. The findings of this study provide theoretical support and technical guidance for the industrial deployment of intelligent ore recognition algorithms.