Date of Award:


Document Type:


Degree Name:

Doctor of Philosophy (PhD)


Computer Science

Committee Chair(s)

Xiaojun Qi


Xiaojun Qi


Haitao Wang


Curtis Dyreson


Vicki Allan


David Brown


Facial Expression Recognition (FER) is the task of predicting a specific facial expression given a facial image. FER has demonstrated remarkable progress due to the advancement of deep learning. Generally, a FER system as a prediction model is built using two sub-modules: 1. Facial image representation model that learns a mapping from the input 2D facial image to a compact feature representation in the embedding space, and 2. A classifier module that maps the learned features to the label space comprising seven labels of neutral, happy, sad, surprise, anger, fear, or disgust. Ultimately, the prediction model aims to predict one of the seven aforementioned labels for the given input image. This process is carried out using a supervised learning algorithm where the model minimizes an objective function that measures the error between the prediction and true label by searching for the best mapping function. Our work is inspired by Deep Metric Learning (DML) approaches to learn an efficient embedding space for the classifier module. DML fundamentally aims to achieve maximal separation in the embedding space by creating compact and well-separated clusters with the capability of feature discrimination. However, conventional DML methods ignore the underlying challenges associated with wild FER datasets, where images exhibit large intra-class variation and inter-class similarity.

First, we tackle the extreme class imbalance that leads to a separation bias toward facial expression classes populated with more data (e.g., happy and neutral) against minority classes (e.g., disgust and fear). To eliminate this bias, we propose a discriminant objective function to optimize the embedding space to enforce inter-class separation of features for both majority and minority classes.

Second, we design an adaptive mechanism to selectively discriminate features in the embedding space to promote generalization to yield a prediction model that classifies unseen images more accurately. We are inspired by the human visual attention model described as the perception of the most salient visual cues in the observed scene. Accordingly, our attentive mechanism adaptively selects important features to discriminate in the DML's objective function.

We conduct experiments on two popular large-scale wild FER datasets (RAF-DB and AffectNet) to show the enhanced discriminative power of our proposed methods compared with several state-of-the-art FER methods.