Mineral processing is the process of separating commercially valuable minerals from their ores. The final quality of the iron ore is tied to the efficiency of its beneficiation process, which can be optimized when the composition of the iron ore is known. Reflected Light Optical Microscopy (RLOM) has been traditionally used to analyze the composition of iron ore samples. However, due to a similar reflectance, the commercially available mounting resins tend to get mixed with the quartz phase. Therefore, it is a well-known problem that, while the major mineralogical phases (mainly hematite goethite and magnetite) can be segmented, it is not possible to segment and analyze the quartz phase by RLOM. Convolutional Neural Networks (CNNs) are a branch of machine learning that have been experiencing a considerable development and thus efficient application in the field of image analysis and classification. As CNNs have been matching and sometimes even outperforming humans, it is reasonable to apply them to this problem, for a human operator can easily distinguish between quartz and resin. After building a databank constituting of 1747 images of resin and 1745 images of quartz for the training set and 442 images of each class for the testing set, the Convolutional Neural Network (CNN) achieved, once trained, success rates above 95%. This success rate is a clear indicator that CNNs can indeed be a solution to this classic problem.