Lithological composition sensor based on digital image feature extraction, genetic selection of features and neural classification

A computer vision system is under development to classify the lithology of rock material on a conveyor belt in a mineral processing plant. The objective of the system is to classify the lithology of the material by considering seven common lithological classes found in the ore: turmaline breccia, other breccias, porphyritic dykes, dacitic diatreme, granodiorites, andesite and riolitic diatreme. The information about the ore lithological composition will help optimize the grinding activity of the plant. A database of 760 digital images of the seven lithological classes was developed. A segmentation procedure was developed to isolate individual rocks. A set of 130 features was extracted from each segmented rock of the database. The genetic algorithm selected 70 of the 130 extracted features with no significant loss in classification performance measured in the test data set. The reduction of the number of inputs also reduced the computation time for feature extraction by nearly 50%