LRA-GNN: Latent Relation-Aware Graph Neural Network with Initial and Dynamic Residual for Facial Age Estimation

TL;DR

This work addresses facial age estimation by modeling non-Euclidean facial structure with a Latent Relation-Aware Graph Neural Network (LRA-GNN). It constructs an initial graph from facial keypoints, enriches it with a random-walk-based global structure, and uses multi-head attention to generate fully connected graphs that capture latent relations, followed by deep residual graph convolutions with adaptive initial and dynamic residuals. A progressive reinforcement learning framework jointly optimizes age-group classification and final regression, using a reward design that accounts for class imbalance and age continuity. Across Morph II, FG-NET, ChaLearn LAP 2016, and UTK-Face, LRA-GNN achieves notable improvements in MAE and CS while maintaining reasonable parameter counts and competitive efficiency, demonstrating robust, generalizable age estimation under varied conditions.

Abstract

Face information is mainly concentrated among facial key points, and frontier research has begun to use graph neural networks to segment faces into patches as nodes to model complex face representations. However, these methods construct node-to-node relations based on similarity thresholds, so there is a problem that some latent relations are missing. These latent relations are crucial for deep semantic representation of face aging. In this novel, we propose a new Latent Relation-Aware Graph Neural Network with Initial and Dynamic Residual (LRA-GNN) to achieve robust and comprehensive facial representation. Specifically, we first construct an initial graph utilizing facial key points as prior knowledge, and then a random walk strategy is employed to the initial graph for obtaining the global structure, both of which together guide the subsequent effective exploration and comprehensive representation. Then LRA-GNN leverages the multi-attention mechanism to capture the latent relations and generates a set of fully connected graphs containing rich facial information and complete structure based on the aforementioned guidance. To avoid over-smoothing issues for deep feature extraction on the fully connected graphs, the deep residual graph convolutional networks are carefully designed, which fuse adaptive initial residuals and dynamic developmental residuals to ensure the consistency and diversity of information. Finally, to improve the estimation accuracy and generalization ability, progressive reinforcement learning is proposed to optimize the ensemble classification regressor. Our proposed framework surpasses the state-of-the-art baselines on several age estimation benchmarks, demonstrating its strength and effectiveness.

Figures (10)

• Figure 1: The illustration of the latent relation (e.g., the dashed part). Existing GNN-based methods utilize the similarity threshold method and may ignore relations between facial key points and wrinkles such as $r_1$, which are not similar enough but important to attain more accurate estimation.
• Figure 2: The Overall Pipeline of our Latent Relation-Aware Graph Neural Network with Initial and Dynamic Residual (LRA-GNN). First, utilizing facial key points as prior knowledge, face images are segmented into patches as graph nodes to construct an initial graph. Then LRA-GNN leverages the multi-attention mechanism to capture the latent relations and generates a set of fully connected graphs. Finally, this set of fully connected graphs is fed into deep residual graph convolutional networks for feature extraction and through progressive reinforcement learning to achieve robust and accurate age estimation.
• Figure 3: The illustration of Random Walk Updating. BFS tends to visit the immediate neighbors of the source node, and DFS tends to explore nodes that are further and further away from the source node. The combination of the two can effectively capture the local and global relations between nodes.
• Figure 4: The illustration of deep feature extraction with the adaptive initial residuals and dynamic developmental residuals. The adaptive initial residuals obtain the personalized characteristics from initial embedding $H^{\left( 0 \right)}$ and hidden embedding $H^{\left( l \right)}$. The dynamic developmental residuals gain the developmental pattern $\alpha ^{\left( l \right)}$ from the residual embedding $\tilde{H}^{\left( l \right)}$.
• Figure 5: The illustration of Progressive RL-based Age Estimation. The age estimation through classification and then regression is defined as a walk-to-the-end problem on a grid.