Romero, M.
, Gutoski, M., Hattori, L.T., Ribeiro, M., Lopes, H.S.
Abstract: Soft biometrics attributes can be useful to perform identification of individuals, since they provide information that can be used to differentiate one individual from another without intrusiveness. Moreover, the large number of surveillance cameras installed in public places allows to acquire videos in real time without much effort. However, this demands an exhaustive process of analysis to be carried out by one or more human observers, which makes necessary the use of methods capable of performing the task automatically. Deep Learning methods have risen in the recent years, achieving state-of-the-art performances for several computer vision tasks such as object recognition, object detection, and image segmentation. This work aims at empirically studying the suitability of a DL approach to perform soft biometrics classication in videos. We evaluate the use of a DL model to learn temporal dependencies, in order to perform soft biometrics classication in videos. For this purpose, we present an approach based on the use of a pre-trained Convolutional Neural Network as feature extractor in combination with a Bidirectional Long Short-Term Memory network to perform the classication.
Keywords: Deep Learning, Convolutional Neural Network, Bidirectional Long Short-Term Memory, Transfer
Learning, Soft Biometrics.
DOI code: 10.21528/lnlm-vol18-no1-art4
PDF file: vol18-no1-art4.pdf
BibTex file: vol18-no1-art4.bib