Identificação de espécies de pássaros em imagens utilizando redes neurais siamesas

Abstract

Bird species classification is essential for ecosystem monitoring and conservation, helping to identify areas at risk for biodiversity and to preserve endangered species. However, the task faces technical challenges due to the large number of species to be classified and the availability of images of lesser-known species. Despite several implementations with machine learning, such as the model in Migliorini (2022), issues such as high computational cost and data imbalance limit the applicability of the models. This work proposes the use of Siamese networks and data augmentation and oversampling techniques to overcome these limitations, seeking to create a more efficient and accurate model for bird species classification. The model was implemented using the Python programming language, with the help of the TensorFlow and Keras libraries, and trained in the Google Colab cloud computing environment. The developed model did not obtain the expected results, reaching 11% accuracy with k=1 and 41% accuracy with k=10 after training.