Deep Learning for automatic classification of mineralogical thin sections

Deep Learning refers to a set of algorithms in machine learning aimed at learning through multiple levels of abstraction. It typically makes use of artificial neural networks characterised by many hidden layers, and has been producing excellent results in image recognition and classification. In this paper, I discuss and compare two different types of Deep Learning architectures: Fully Connected and Convolutional Residual Networks. Using a set of test images, I show that these neural networks can be applied, with different effectiveness, for classifying complex images extracted from mineralogical thin sections. Both techniques produce reliable results, although the learners are trained on a limited number of examples. However, in the case of Fully Connected Networks, the vanishing gradient problem represents a crucial limitation when increasing the number of hidden layers. In fact, the classification results remain unchanged, or even degrade, when additional neural layers are progressively stacked. Instead, the same problem is overcome in Convolutional Residual Networks, through a technique based on shortcut connections. Hundreds of hidden layers are used for improving the classification performances, without incurring into the degradation of learning capabilities.