Abstract:
The task of modeling and identifying people’s emotions using facial cues is a complex problem in computer vision. Normally
we approach these issues by identifying Action Units (AUs), which have many applications in Human Computer Interaction (HCI).
Although Deep Learning approaches have demonstrated a high level of performance in recognizing AUs and emotions, they require
large datasets of expert-labelled examples. In this article, we demonstrate that good deep features can be learnt in an unsupervised
fashion using Deep Convolutional Generative Adversarial Networks (DCGANs), allowing for a supervised classifier to be learned from
a smaller labelled dataset. The paper primarily focuses on two key aspects: firstly, the generation of facial expression images across
a wide range of poses (including frontal, multi-view, and unconstrained environments), and secondly, the analysis and classification
of emotion categories and Action Units. We demonstrate an enhanced ability to generalize and achieve successful results by using a
different methodology and a variety of datasets for feature extraction and classification. Our method has been thoroughly tested through
multiple experiments on different databases, leading to promising results.