Marble Quality Assessment with Deep Learning Regression

Abstract

Natural rock tile classification, with the use of computer vision and machine learning techniques, is a methodology well documented in academic literature. The broad variety of textures present on the rock tiles’ surface, creates many ornamental patterns. This leads to rather different-looking surfaces being classified into the same group based on certain criteria. In this paper, regression is tested on dolomitic marble tiles to assign a quality value (QV) ranging from zero to one. Five Convolutional Neural Networks were tested on a dataset containing digital images (DI) of natural rock tiles and their QV assigned by an expert. The dolomitic samples were provided by Solakis Marble Industries S.A. The DIs were acquired by an automatic machine developed by Intermek S.A. MobileNetV2 (MNV2) achieved the best result with a Mean Absolute Percentage Error (MAPE) of 31.20% to the actual QV assigned to the marble tile.