Gemstone classification using ConvNet with transfer learning and fine-tuning
Willian M. Freire, Aline Maria Malachini Miotto Amaral, Yandre M. G. Costa
Abstract
A gemstone is a piece of mineral crystal that, when cut and polished, is used to make jewelry or adornments. Identifying a gemstone is a complex task due to several aspects related to its nature. Deep Learning (DL) is a research field that enables the construction of systems capable of acquiring knowledge. Thus, this work proposes to use an DL model to classify gemstones. Additional strategies such as data augmentation (DA), image clipping, and transfer learning (TL) techniques have also been experimented with. For this last technique, the InceptionV3 model was compared with other models already present in the literature. They were compared based on accuracy and loss metrics, and in cases where F1, Precision, and Recall metrics were available, the evaluation was also performed. Experiments were conducted on a dataset composed of gemstones categorized into 87 classes. The proposed model achieved an accuracy of 72%, corroborating the feasibility of automating this task.