Generación de un modelo de reconocimiento facial en videos usando redes neuronales convolucionales bajo un esquema de aprendizaje por transferencia

Abstract

The current use of CNN convolutional neural networks in machine learning has given way to countless activities and processes that, until a few years ago, were unthinkable that can be carried out by computer programs.

Taking advantage of these important advances, this work uses the transfer learning technique to be able to use convolutional neural networks already trained in facial recognition. In this way, taking into account the great accuracy presented by the convolutional neural network VggFace, this network was used as a basis to be able to carry out the learning transfer technique. Specifically, a database of images of faces collected for this purpose was created. From this basis, a facial recognition model was trained using the VggFace neural network. With the trained model, predictions were made of the frames that make up videos in which multiple people appear. The predictions on the frames of these videos were subjected to a temporal coherence algorithm.

Experiments show that the prediction accuracy of the model increases markedly with the use of temporal coherence. This is accomplished by sliding a window into the predictions made on the frames that make up multiple people's videos.