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

Age verification using facial images enhances security, online child safety, and access control, though existing classification models often struggle with generalization due to homogenous racial datasets and feature evaluation. To improve diversity, we selected samples from four benchmark datasets: UTK-Face, Fg-Net, Morph, and All-Age-Faces. We explored the predictive potential of local, global, and hybrid facial features, extracting local features with two Local Binary Pattern (LBP) models—one with 16,384 features based on uniform patterns and one with 10 features from fixed-length histogram bins. For global features, we applied a geometric ratio model yielding 12 features. Hybrid feature sets combined histogram-based LBP and geometric ratios or used a Hybrid Partial Active Appearance Model (HPAAM) with 10,136 features. Feature predictiveness was assessed with Chi-Square, ANOVA, and Information Gain tests, followed by classification experiments, both balanced and unbalanced via K-Means. Results indicate that histogram-based LBP and geometric ratios consistently outperform uniform LBP and HPAAM, achieving average accuracies of 78.5% and 76.5%, respectively, and up to 80% on balanced data. Combining histogram-based LBP and geometric ratios further improved accuracy to 83%. These findings suggest that conservative feature extraction with fewer features enhances accuracy, reduces the need for extensive feature selection, and lowers computational demands by mitigating noise from irrelevant features.