Création d’images de matériaux granulaires avec la plateforme de développement de jeux Unity et prédiction de la granulométrie à l’aide de textures et de réseaux de neurones

Abstract

Image analysis techniques can be used to determine the particle size distribution of granular materials from images. The most recent methods for grain size analyses using image analysis techniques use textural parameters as inputs for artificial neural networks. The performance of these methods depends in part on the quality and quantity of the images used for training the neural networks. There is currently a significant gap in the available datasets in the literature. The first objective of this thesis is to create a database of realistic images of granular materials with the Unity game engine. The images were created so that the particles have different shapes and sizes. These images had to respect a series of imposed particle size distributions, and had to appear as natural as possible. A total of 3003 images were prepared with particle sizes ranging from 75 to 1180 μm. The second objective of this thesis is to verify the influence of the dataset on particle size predictions using textural parameters. To do so, two additional databases were prepared: Yade and Unity+Yade. The Yade database includes 3003 randomly selected images from the Pirnia et al. (2018) database. The images in this database show spherical particles created using the discrete element method. The Unity+Yade database combines the Yade and Unity images for a total of 6006 images. The images of the three databases have the same size, scale, and particle mass. Three sets of textural parameters were extracted from each image: the mean and standard deviation of local entropy, the mean and standard deviation of Haar wavelet transforms, and Haralick textures. These characteristics were used as inputs for neural networks. The networks were trained to predict the percentage passing for sieves of 106, 150, 250, 425, and 710 μm. The neural network performance was evaluated with the root mean square error (RMSE) of the percentage passing. For the three sets of textural parameters, the mean RMSE is 7.0% for the Yade images, 8.4% for the Unity images and 8.5% for the Unity+Yade dataset. For the three datasets, local entropy performed better (7.3%) than Haar wavelet transforms (7.9%) and Haralick textures (8.7%). These results are slightly worse than those of Manashti et al. (2020) who worked with 53130 Yade images and obtained an RMSE between 5.6% and 7.0% for the same three textural parameters.

Date	4 Dec 2020
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	François Duhaime (Supervisor) & Matthew Toews (Co-supervisor)