Constellation, le dépôt institutionnel de l'Université du Québec à Chicoutimi

Résumé

Identifying and counting individual mineral grainsc composing sand is an important component of many studies in environment, engineering, mineral exploration, ore processing and the foundation of geometallurgy. Typically, silt (32–128 μm) and sand (128–1000 μm) sized grains will be characterized under an optical microscope or a scanning electron microscope. In both cases, it is a tedious and costly process. Therefore, in this paper, we introduce an original computational approach in order to automate mineral grains recognition from numerical images obtained with a simple optical microscope. To the best of our knowledge, it is the first time that the current computer vision based on machine learning algorithms is tested for the automated recognition of such mineral grains. In more details, this work uses the simple linear iterative clustering segmentation to generate superpixels and many of them allow isolating sand grains, which is not possible with classical segmentation methods. Also, the approach has been tested using convolutional neural networks (CNNs). However, CNNs did not give as good results as the superpixels method. The superpixels are also exploited to extract features related to a sand grain. These image characteristics form the raw dataset. Prior to proceed with the classification, a data cleaning stage is necessary to get a usable dataset for machine learning algorithms. In addition, we present a comparison of performances of several algorithms. The overall obtained results are approximately 90% and demonstrate the concept of mineral recognition from a sample of sand grains provided by a numerical image.