Table 3. Comparison of facial age estimation performance on the FG-NET dataset using the LOPO cross-validation protocol. Reported MAE values are extracted from existing studies that explicitly disclose LOPO-based evaluation on FG-NET. The final row presents the result of the proposed method using MobileNetV2 with randomly initialized weights and the CAE-based data augmentation strategy, which achieves the lowest MAE among all listed approaches.

Method	MAE
Aging Pattern Subspace (AGES) Algorithm [45]	6.22
Relevance Vector Machine (RVM) [46]	6.2
Regression with Uncertain Nonnegative Labels [47]	5.78
Improved Iterative Scaling (ISS) Algorithm [48]	5.77
Ranking with Uncertain Labels [49]	5.33
Synchronized Submanifold Embedding (SSE) [50]	5.21
LBP Kernel Density Estimate [51]	5.09
Locally Adjusted Robust Regressor (LARR) [52]	5.07
Probabilistic Fusion Approach (PFA) [53]	4.97
Ranking-KNN [54]	4.97
Rank-based Age Value Estimation [55]	4.89
Ordinal Discriminative Features (PLO) [56]	4.82
Biologically Inspired Features (BIF) [57]	4.77
Deep Expectation (DEX) [58]	4.63
Component and Holistic BIF [59]	4.6
Ordinal Hyperplane Ranking (OHRank) [60]	4.48
Biologically Inspired Active Appearance Model [61]	4.18
Mean-Variance Loss [62]	4.1
Deep Regression Forests (DRFs) [24]	3.85
Deep Hybrid-Aligned Architecture (DHAA) [63]	3.72
Adaptive Mean-Residue Loss [64]	3.61
Extended BIF (EBIF) [65]	3.17
Deep Random Forests [66]	3.05
MobileNetV2 (Random Init.)+Proposed CAE-DA	2.77