Adaptive Latent Distribution Margin GAN for Imbalanced Data Classification

Facial Expression Recognition (FER) has achieved significant development as a result of the development of deep learning. However, the constrained size of FER datasets constrains the generalization capability of expression recognition models, resulting in suboptimal model performance. In addition, the currently available training datasets for FER suffer from imbalances that give preference to certain majority classes, resulting in diminished test accuracy specifically for the minority classes. In this paper, we propose a Adaptive Latent Distribution Margin Generative Adversarial Network (ALDMGAN). It employs a distinct latent distribution model for each class, facilitating a clear demarcation between minority and majority classes. In addition, we propose a objective function that minimizes the distribution disparity between classes, resulting in a closer proximity of their latent distributions. Extensive experiments have demonstrated that ALDMGAN substantially enhances the classification accuracy of unbalanced data.