Publication

Abstract : Facial emotion recognition (FER) remains challenging under limited data, noise, and occlusion. This study introduces an Actor–Critic Convolutional Deep Belief Network (ACCDBN) that unifies Generative Adversarial Network (GAN)–based augmentation, deep probabilistic feature learning, and reinforcement-driven optimization. Conditional GANs expand minority emotion classes, enhancing data diversity, while the CDBN extracts hierarchical texture features through convolutional and restricted Boltzmann layers. An Actor–Critic module dynamically refines representations by rewarding accurate emotion classification and penalizing uncertain predictions. Trained and validated on the CK+ dataset with five-fold cross-validation, the proposed model achieves higher accuracy and stability than CNN, LSTM, and ResNet-50 baselines, maintaining strong performance under noise and occlusion. The approach demonstrates how reinforcement-guided generative learning can improve both accuracy and robustness in FER tasks. 1. To implement this, the research utilised the publicly available Cohn-Kanade+ dataset, consisting of eight classes with samples of 920 grey-scale images. 2. An improved ACCDBN model outperformed with 90.4% accuracy and 0.69 MCC (Mathew’s Correlation Coefficient) in 5-fold cross-validation using the cGAN-generated dataset and 87% on the CK+ dataset 3. The main objective is to present an advanced facial emotion recognition (FER) system that combines a Convolution Deep Belief Network (CDBN) with a model-free reinforcement learning technique, namely the actor-critic approach.