Very deep convolutional neural networks for face identification
Document typeMaster thesis
Rights accessOpen Access
The goal of this thesis is to evaluate the face identification problem using very deep convolutional neural networks. In recent years, the use of CNN, with a large amount of images in databases, have made the deep learning technique very performant. The problems in training a network from scratch, such as having sufficient hardware resources and large databases, can be overcome using the finetune technique on pretrained models. This thesis evaluate the performance in finetuning for face classification the most recent CNN architectures which have obtained the best results at ImageNet Large-Scale Visual Recognition Challenge (ILSVRC) in the last years, in particular VGG, GoogLeNet and ResNet. All the pre-trained models of the CNNs were downloaded from the MatConvNet website. VGG-16 has shown best results in face classification which was followed with ResNet-101 and GoogLeNet that are the matter of this thesis.
In very recent years, several classification problems in computer vision, have boosted its performance by using Deep Learning techniques, in particular Convolutional Neural Networks (CNNs). The topic of the research project will focus in exploring state of the art deep learning architectures in computer vision applications. Recently architectures like GoogleNet and VGG have shown to perform better than other architectures.