Do Deep Learning Models Generalize Facial Emotion Recognition in Different Age Groups?

The prevalence of images of adults in facial emotion recognition (FER) databases is an evident bias in emotion classification. Although there are sets of FER images exclusively of children or adults aged 60 and over, the capture scenario is generally controlled and has staged emotions. Many studies in this area focus on increasing task performance by using a single data collection for training and testing; however, due to the diversity of individuals portrayed and photographs collected, this strategy generally fails when applied to new images or in real-world circumstances. In this paper, we assess the generalization potential of deep learning models that have been pre-trained in general-purpose databases such as ImageNet and refined using the databases CK+, DEFSS, FACES (60+), MUG, and NIMH-ChEFS with cross-databases in tests. Through this study, which assesses six different datasets involving three distinct age groups through cross-database evaluation using a clear experimental protocol, we were able to verify that models fine-tuned with data sets composed exclusively of children's images have accuracy above 80%, reaching 88% when used to predict images of children; models fine-tuned with images of young adults are more accurate in classifying young adults than children or adults aged 60 and over, and the dataset of adults aged 60 and over in FACES dataset demonstrated accuracy above 80% in predicting images from other datasets and 96% in predicting images within its own test set.