Comparing Convolutional Neural Networks and Deep Metric Learning Methods for Classification of Export Watermelon (Citrullus lanatus) Varieties

Abstract

This study uses computer vision methods to understand and classify export watermelon (Citrillus lanatus) varieties. A data set comprised of 330 images from three wa-termelon varieties (Joya, Anna, Quetzali) was constructed. This study compares using two distinct methods for this task. First, convolutional neural network (CNN) with transfer learning (TL) pretrained on VGG19, ResNet50 and EfficientNetB0 networks. Also, Deep Metric Learning (DML) methods in a triplet neural network architectures. Results suggest that a slight overfitting can be found due to both methods showing that the Joya variety is not easily classifiable. In conclusion both strategies demonstrate to be feasible for the task, 84% for TL and 81% for DML and specifically. These methodologies can be in the future updated to encompass Generative Adversarial Networks or n-shot learning. Finally they can be integrated into a watermelon detection product in order to help local and global producers.