Meta-RL Based Micro-Expression Recognition Framework Using MAML with Calibrated Regression Function

Abstract Micro-Expression Recognition (MER) plays a crucial role in understanding human emotions, yet its effectiveness is often hindered by the transient and subtle nature of micro-expressions. This article presents a novel MER framework integrating Calibrated Regression with Maximum Mean Discrepancy (MMD), Model-Agnostic Meta-Learning (MAML), and Meta-Reinforcement Learning (Meta-RL) to enhance recognition accuracy and adaptability. The 2D Convolutional Neural Network (2DCNN) is employed as the backbone for feature extraction, capturing fine-grained spatial details of Facial Expressions (FEs). To address challenges in feature alignment, a Heteroscedastic Neural Network (HNN) is introduced for predictive uncertainty estimation. A two-stage learning process is applied, where Negative Log-Likelihood (NLL) optimization refines model parameters, and MMD ensures better alignment between micro- and macro-expressions. Additionally, the Meta-RL framework optimizes feature learning through characterizing the optimal gap of the stationary points achieved using MAML and improving generalization. Extensive experiments on benchmark datasets like SAMM and CASME II shows the advantage of the introduced approach, achieving 96.74% accuracy on SAMM and 98.84% accuracy on CASME II, surpassing state-of-the-art models. The results highlight the model’s robustness, adaptability, and effectiveness, making it well-suited for real-world micro-expression analysis applications.