Clasificación de minerales usando procesamiento de imágenes digitales

Abstract

In mining, sensor-based ore sorting is a technology used to classify different types of minerals in order to process them using the optimal method, thus, reducing operation costs. This thesis presents a mineral classification algorithm based on image processing and machine learning used to classify minerals from a gold and silver mine with better performance than commercial ore sorting machines. The algorithm comprises three main steps: image segmentation based on binary thresholds, feature extraction using color statistics, principal component analysis and wavelet texture analysis, and classification using neural networks. The algorithm was trained using color images from 156 rocks that belong to 4 types of minerals, which an expert geologist manually classified. Then, the algorithm was tested using 46 rocks, which were analyzed in a laboratory to determine their gold and silver grades. The method was validated by measuring the processing time for each rock, and the classification accuracy of each image from the test set according to the geologist and the chemical assays. Additional tests were performed to compare the proposed method with other classification algorithms, such as convolutional neural networks and support vector machines. The proposed method was the most accurate in the tests that used the geologist’s classification as labels. However, in the tests that used the chemical composition of the rocks as labels, the performance of the SVM model was comparable to the proposed method, even exceeding the F1-score of the proposed method by 0.4 % in some of the tests. To summarize, the maximum processing time of the proposed method was 44ms per image, and its accuracy was 97.6 %, which is higher than the 95 % required by the mining company based on previous tests with commercial ore sorting machines.

Date of Award	2021
Original language	Spanish (Peru)
Supervisor	Víctor Manuel Murray Herrera (Asesor)