Exploiting Deep Metric Learning for Mable Quality Assessment with Small and Imbalanced Image Data

Abstract

The classification of ornamental dolomitic marble stone tiles has been an issue in the past years, even more so according to their aesthetical criteria. Quality control and product classification during the final stage of a production line is the main problem of this step, which, when done right, can increase profitability. Machine Learning has been employed in many cases to improve and accelerate the decision and assessment process of this step. Due to the unique nature of the problem, the image datasets constructed can be heavily unbalanced, as there is no control over the number of marble tiles that are collected for each class. This paper examines the application of metric learning and more specifically Siamese networks, for the classification of dolomitic marble tiles, examining the performance of 7 convolutional neural networks as feature extractors. The results are then compared to the application of transfer learning techniques on the same convolutional networks. The experiments conducted revealed the high robustness of the metric learning approach, by providing very low standard deviation (stdev 0.53%) between the models' performance, compared to transfer learning where results per model vary (stdev 2.53 %) to a higher degree.