| Article ID | Journal | Published Year | Pages | File Type |
|---|---|---|---|---|
| 214169 | International Journal of Mineral Processing | 2011 | 9 Pages |
In mining, rock classification plays a crucial role at different stages of the extraction process ranging from the design of the mine to mineral grading and plant control. In this paper we present a new method to improve rock classification using digital image analysis, feature selection based on mutual information and a voting process to take into account boundary information. We extract rock color and texture features and using mutual information we selected 14 from 36 features to represent the data in a lower dimensional space. The original image was divided into sub-images that are assigned to one class based on the selected color and texture features using a set of classifiers in cascade. Additionally, using rock boundary information, a voting process for the sub-images within the same blob is performed. We compare our results based on sub-image classification to those obtained after the voting process and to those previously published on the same rock image database. We show that the RMSE on rock composition classification on a test database decreased 8.8% by using our proposed voting method with the automatic segmentation with respect to direct sub-image classification. The RMSE decreased 29.5% relative to previously published results with the same database using a mixture of dry and wet rock images. The RMSE decreased even more if we considered separately dry and wet rocks. Our proposed method could be implemented in real-time to estimate mineral composition and can be used for online ore sorting and/or classification.
► We use mutual information to select rock features improving composition estimation. ► We propose a voting method among sub-images to improve rock classification. ► Rock composition estimation improves by feature selection and sub-image voting. ► Our results are compared to those previously published on the same rock database. ► Computational time can be reduced by our proposed feature selection method.
