Résumé:
Since the emergence of deep learning approaches, facial expression recognition has
grown significantly, and these systems have become increasingly more efficient. However,
effective learning requires large databases which are not always available.
To address this, databases need to be augmented using appropriate data augmentation
techniques (DA), which are typically based on geometric transformations or oversampling
augmentations. These techniques allow to increase the size of the training database and
to rebalance it.
In this project, we propose to explore and evaluate the use of three data augmentation
techniques, as well as the effect of their hybridization, on a facial expression recognition
system. Two deep learning models have been used for this evaluation, the VGG16 and the
Resnet50. The augmentation techniques tested are geometric transformations, SMOTE
and Cycle-Gan.
The results obtained show that the operation of the RMF systems requires a large volume
of data for learning and that GANs offer an interesting alternative to overcome this data
deficit.