Training Convolutional Neural Networks with Imbalanced Datasets: an Empirical Evaluation on the WildCat Dataset

This thesis aims to exploit deep learning and computer vision techniques in wildlife monitoring. In zoology, wildlife monitoring is a crucial task for keeping track of species movement patterns and demographics. One technique to track wild animals consists of capturing images with spatio-temporal tags to later recover species movements in space and time. Such images are usually acquired through camera traps in the wild or from veterinarians or zoologists when wild animals get injured or found dead. This thesis focuses on the latter case, and uses convolutional neural networks (CNN) to automatically classify three categories of cats: European Wildcat, Domestic Cats and Hybrid Cats. The architectures employed here are Deep Residual Learning Architecture (Resnet 18/50) and Very Deep Convolutional Networks Architecture (VGG16). The aforementioned CNNs models are trained with a dataset of images, named wildcat dataset, collected from various researchers around the globe and made available to me during the internship at Fuorisentiero. Despite image classification being a task in which we saw significant advances thanks to deep convolutional architecture, classes imbalance is still a problem that considerably affects their performances. The wildcat dataset belongs to the category of imbalanced sets, with a class distribution strongly biased toward the wildcat class. In this work, we explored different methodologies to deal with datasets' classes imbalances in the context of CNN. We considered, analyzed and evaluated the following methods: Over/Under-sampling and Cost Sensitive Loss. We further report other methods in the literature, such as the Synthetic Minority Oversampling Technique(SMOTE). The experiments showed that such methods are of primary importance for CNN's performance, being able to considerably improve their accuracy.