Reconstructing Neutral Face Expressions with Disentangled Variational Autoencoder

This study tackles unsupervised learning for disentangled representations in facial expression generation. It introduces a novel deep architecture by combining FactorVAE and \(\beta \) -VAE concepts, incorporating the Ranger optimizer and Dropout layers. The goal is to learn disentangled representations that capture essential facial expression factors, like the neutral expression, while ensuring independence between different expression dimensions. This approach achieves faster convergence and improved optimization, striking a better balance between disentanglement and reconstruction quality. The method enables accurate and diverse facial expression generation and enhances model generalization through adaptive learning rate adjustments. The learned disentangled representations also facilitate the generation of realistic and interpretable facial expressions, making it a promising approach for various facial expression generation tasks.