Autor(es)

Benato, Barbara C. ; Gomes, Jancarlo F. ; Telea, Alexandru C. ; Falcao, Alexandre Xavier

• Tipo de documento: Artigo Científico
• Fonte: PROGRESS IN PATTERN RECOGNITION, IMAGE ANALYSIS, COMPUTER VISION, AND APPLICATIONS, CIARP 2021; v. 12702, p. 11-pg., 2021-01-01.

Resumo

Expert human supervision of the large labeled training sets needed by convolutional neural networks is expensive. To obtain sufficient labeled samples to train a model, one can propagate labels from a small set of supervised samples to a large unsupervised set. Yet, such methods need many supervised samples for validation. We present a method that iteratively trains a deep neural network (VGG-16) from labeled samples created by projecting the features of VGG-16's last max-pooling layer in 2D with t-SNE and propagating labels with the Optimum-Path Forest semi-supervised classifier. As the labeled set improves along iterations, it improves the network's features. We show how this significantly improves classification results on test data (using only 1% to 5% of supervised samples) of three private challenging datasets and two public ones. (AU)

• Processo FAPESP: 19/10705-8 - Aprendizado Ativo Visual guiado por Projeções de Características
• Beneficiário: Bárbara Caroline Benato
• Modalidade de apoio: Bolsas no Brasil - Doutorado
• Processo FAPESP: 14/12236-1 - AnImaLS: Anotação de Imagem em Larga Escala: o que máquinas e especialistas podem aprender interagindo?
• Beneficiário: Alexandre Xavier Falcão
• Modalidade de apoio: Auxílio à Pesquisa - Temático