Brazilian Birds of Prey - A New Dataset and Classification with Deep Neural Networks

Abstract

Birds of prey play an essential role in maintaining the health of their ecosystems. In Brazil, where there is a vast amount of biodiversity, the identification and monitoring of predatory birds are essential for maintaining the ecosystem. However, developing computational methods for the classification of predatory birds based on images is not trivial, given the many possible variations, such as angles, lighting, birds camouflage, and others. Nowadays, Transfer Learning (TL) approaches have gained popularity for many applications due to a large amount of knowledge previously acquired by models from huge datasets, which can be leveraged for other similar problems. In this paper, we present a dataset of birds of prey images and also introduce a baseline classification benchmark using the TL approach. The experiments were divided into two subcategories: families and species classification. The proposed dataset contains 42,475 samples, from 6 families and 41 species. The samples of the dataset contain birds in different positions and angles, with great variety with respect to background and illumination. Baseline results achieved an F1-Score of 92% in family and 80% in species classification.