A Data Augmentation Methodology to Improve Age Estimation Using Convolutional Neural Networks
Ítalo de Pontes Oliveira, Joao Lucas Peixoto Medeiros, Vinicius Fernandes de Sousa, Adalberto Gomes Teixeira, Eanes Torres Pereira, Herman Martins Gomes · 2016
Recent advances in deep learning methodologies are enabling the construction of more accurate classifiers. However, existing labeled face datasets are limited in size, which prevents CNN models from reaching their full generalization capabilities. A variety of techniques to generate new training samples based on data augmentation have been proposed, but the great majority is limited to very simple transformations. The approach proposed in this paper takes into account intrinsic information about human faces in order to generate an augmented dataset that is used to train a CNN, by creating photo-realistic smooth face variations based on Active Appearance Models optimized for human faces. An experimental evaluation taking CNN models trained with original and augmented versions of the MORPH face dataset allowed an increase of 10% in the F-Score and yielded Receiver Operating Characteristic curves that outperformed state-of-the-art work in the literature.