Classificação de imagens de bovinos utilizando redes neurais convolucionais

Abstract

The livestock activities increasingly rely on information technologies that can contribute in several steps of the productive chain of the sector. In particular, many problems of Computer Vision focused on the livestock sector have been addressed through deep learning techniques and the results achieved have been satisfactory, demonstrating the potential that such techniques have to benefit these activities. Furthermore, approaches like this can be very advantageous, since they are non-invasive applications that can generate results in real time. Thus, this work proposes the development of three types of image classification models generated from Convolutional Neural Networks (CNNs): category, pose, and breed. A series of experiments were performed, of which the best results are reported in this paper. From a sequence of experiments and evaluations, it was possible to achieve satisfactory results for the cases addressed. More specifically, the pose and category classifier models achieved accuracies greater than 90%, while the models generated for the breed problem achieved worse results (around 83%), which demonstrated that it is a more challenging problem. Analyzing all the results, we verified the potential that such approaches have to achieve even better results in future works.