Identificação de aves da fauna do Brasil através de aprendizado profundo para classificação de imagens

Abstract

Deep learning use have become frequent in several areas of biology. Specifically, Convolutional Neural Networks (CNNs) are widely used for image prediction and have drawn the attention of biologists since they can automate analyzes that by conventional methods would require great effort and time to perform, such as species identification. In this study, we used transfer learning to obtain a classifier of bird species from the Brazilian fauna. Using a image database with 1953 classes and more than 600000 images, we developed our classifier model using the ResNet50V2 architecture. The model reached 45% of accuracy in the validation set, a performance that can be improved in future studies with the exploration of other techniques and methods. This work highlights the importance of providing to society a tool to identify birds from one of the greatest biodiversity on Earth.