Multifeature Fusion for Facial Expression Recognition

This paper introduces a novel approach for FER utilizing multifeature fusion based on supervised learning, leveraging the inherent strengths of LBP, HOG, and SIFT descriptors, a combinatorial technique was employed to discern the most efficient fusion methodology. Two fusion strategies were assessed: direct concatenation and the Z-score method, with the latter demonstrating superior computational efficacy. Normalization was introduced, to enhance descriptor performance. Subject-dependent (SD) and subject-independent (SI) were employed; comparative evaluations showcased the accuracy of our approach’s (100%) proficiency on JAFFE, CK+, and FER2013 datasets compared to recent works (99.2%) (Comput Intell Neurosci 2021:1–10). Despite the notable outcomes, the system’s dependency on well-labeled datasets was a limitation. Consequently, future research avenues are suggested to incorporate poorly labeled and unlabeled datasets, facilitating the exploration of less structured datasets. This study paves the way for advancing FER, underscoring the potential of multifeature fusion and the importance of rigorous feature normalization.