Unveiling the Human-like Similarities of Computerized Facial Expression Recognition: An Empirical Exploration via Explainable AI
Authors: F. Xavier Gaya-Morey, Silvia Ramis-Guarinos, Cristina Manresa-Yee, Jose M. Buades-Rubio
Summary: Facial features recognition is important for human habits evaluation, and deep studying has enabled fashions that may outperform people. Nonetheless, it’s unclear how intently they mimic human processing. This research goals to discover the similarity between deep neural networks and human notion by evaluating twelve completely different networks, together with each common object classifiers and FER-specific fashions. We make use of an revolutionary international explainable AI methodology to generate heatmaps, revealing essential facial areas for the twelve networks educated on six facial expressions. We assess these outcomes each quantitatively and qualitatively, evaluating them to floor fact masks based mostly on Friesen and Ekman’s description and amongst them. We use Intersection over Union (IoU) and normalized correlation coefficients for comparisons. We generate 72 heatmaps to focus on important areas for every expression and structure. Qualitatively, fashions with pre-trained weights present extra similarity in heatmaps in comparison with these with out pre-training. Particularly, eye and nostril areas affect sure facial expressions, whereas the mouth is constantly necessary throughout all fashions and expressions. Quantitatively, we discover low common IoU values (avg. 0.2702) throughout all expressions and architectures. One of the best-performing structure averages 0.3269, whereas the worst-performing one averages 0.2066. Dendrograms, constructed with the normalized correlation coefficient, reveal two major clusters for many expressions: fashions with pre-training and fashions with out pre-training. Findings counsel restricted alignment between human and AI facial features recognition, with community architectures influencing the similarity, as comparable architectures prioritize comparable facial areas
