A computational study on aging effect for facial expression recognition

Abstract

This work uses newly introduced variations of the sparse representation-based classifier (SRC) to challenge the issue of automatic facial expression recognition (FER) with faces belonging to a wide span of ages. Since facial expression is one of the most powerful and immediate ways to disclose individuals' emotions and intentions, the study of emotional traits is an active research topic both in psychology and in engineering fields. To date, automatic FER systems work well with frontal and clean faces, but disturbance factors can dramatically decrease their performance. Aging is a critical disruption element, which is present in any real-world situation and which can finally be considered thanks to the recent introduction of new databases storing expressions over a lifespan. This study addresses the FER with aging challenge using sparse coding (SC) that represents the input signal as the linear combination of the columns of a dictionary. Dictionary learning (DL) is a subfield of SC that aims to learn from the training samples the best space capable of representing the query image. Focusing on one of the main challenges of SC, this work compares the performance of recently introduced DL algorithms. We run both a mixed-age experiment, where all faces are mixed, and a within-age experiment, where faces of young, middle-aged, and old actors are processed independently. We first work with the entire face and then we improve our initial performance using only discriminative patches of the face. Experimental results provide a fair comparison between the two recently developed DL techniques. Finally, the same algorithms are also tested on a database of expressive faces without the aging disturbance element, so as to evaluate DL algorithms' performance strictly on FER.