Data Augmentation and Transfer Learning Applied to Charcoal Image Classification

This paper proposes the evaluation of data augmentation impact in the process of microscopic charcoal image classification. Two data augmentation approaches were explored, namely morphological transformations and sub-images. From the augmented data, a pre-trained Inception-v3 network was used to train a classifier of charcoal species. The best result was found through the technique of sub-images, with an average accuracy of 99.36%.