An Evaluation of Features Extracted from Facial Images in the Context of Binary Age Classification

Age verification via facial images is used to enhance security, ensure online child safety, and manage access control with many approaches using classification models. However, many existing classification approaches face generalisation challenges due to homogenous racial nature of datasets. Moreover, many approaches fail to explicitly address the predictive potential of extracted facial features. To address the lack of racial diversity we selected representative samples from four different benchmark datasets: UTK-Face, Fg-Net, Morph and All-Age-Faces. Subsequently, we examined the predictive potential of local, global and hybrid sets of facial features. We extracted local features using two types of Local Binary Pattern (LBP)—one based on the number of uniform patterns resulting in 16384 features and one with 10 features based on fixed length histogram. To extract global features, we used a geometric ratio model with 12 features. Finally, two hybrid feature sets were generated using hybrid partial active appearance model(HPAAM) with 10136 features and a model combining the 10 bin LBP features and the 12 geometric ratios. We then assessed feature predictiveness for larger feature sets via Chi-Square, Anova and Information Gain. Finally, we conducted classification experiments with and without data K-Means balancing to evaluate the accuracy of the models generated based on various feature sets. The results show that histogrambased LBP and geometric ratios models consistently outperform the uniform LBP and HPAAM, even when balancing and feature selection were employed: uniform LBP and hybrid partial AAM lead to average accuracies of 70.48% and 72.68%, respectively. Conversely, the histogram-based LBP and facial ratio models lead to average accuracies of 78.5% and 76.5%, respectively, reaching 80% on balanced data. We observed further improvements in accuracy, precision and recall when histogrambased LBP and facial ratios features were combined reaching 83% accuracy. These findings indicate that using large subsets of features does not lead to better performance due to irrelevant features that introduce noise. A more effective approach is to use conservative feature extraction techniques with smaller number of features, which not only improves model accuracy and reduces the need to apply feature selection, but also assumes a reduced computational overhead.