Lithological classification and chemical component estimation based on the visual features of crushed rock samples

Lithological classification and monitoring of ore quality are challenging issues in mine operation, especially to maintain desired feed for processing minerals plants. Quantification of visual features is an innovative method to analyze rock components. In this paper, an analysis of vision-based rock type and classification algorithm is proposed for fast, inexpensive, and reliable identification process compared with common chemical analysis. To evaluate the proposed algorithm, samples were collected from the Novin limestone mine in Iran. Based on chemical and thin-section studies, the samples were classified into three different groups as calcium carbonate, dolomite, and dolomitic limestone. The limestone samples were crushed, sieved, and, respectively, divided into three size fractions as 2.5–7, 1.5–2.5, and 0.1–1.5 cm. A set of 58 images as large and medium size samples and 78 images as fine size samples were taken in the laboratory environment. Features were extracted from the captured images and reduced by applying principal component analysis (PCA). The support vector machine (SVM) and Bayesian techniques were used for classification. The best classification accuracy was about 80 and 90% in limestone and dolomite rock samples, respectively. Then, a multi-layer perceptron (MLP) neural network was employed to predict chemical compositions percentages. The determination coefficient within the range of 0.76 to 0.85 was observed and predicted values confirmed good performance of the grade estimation. The outputs illustrated that the proposed intelligent and automated technique can be successfully applied to monitor ore grade and classify lithology in different stages of mining projects.