Extreme Sparse Learning for Robust and Comprehensive Facial Emotion Recognition

Abstract

Recognizing natural emotions from human faces is a fascinating area with diverse applications, including human-computer interaction, automated tutoring systems, multimedia retrieval, intelligent environments, and driver assistance systems. Traditionally, facial emotion recognition systems are tested on controlled laboratory datasets, which fail to reflect the challenges of real-world environments. To address this, the paper introduces an approach called extreme sparse learning, which simultaneously learns a dictionary (basis set) and a nonlinear classification model. This approach integrates the discriminative capabilities of extreme learning machines with the reconstruction strengths of sparse representation, ensuring robust classification even with noisy and imperfect data from natural settings. Furthermore, a novel local spatio-temporal descriptor, designed to be both distinctive and pose-invariant, is proposed. The framework achieves state-of-the-art recognition accuracy on both acted and spontaneous facial emotion datasets.