DCGAN Based Facial Expression Synthesis with Feature Extraction and Cluster Grouping

This research proposes a novel variant of Deep Convolutional Generative Adversarial Network (DCGAN) model for realistic facial expression image generation. It incorporates embedding and reshape layers on both Generator and Discriminator models in conjunction with a novel cluster grouping technique to improve image generation performance. The model is subsequently used to generate images with seven facial expressions, i.e. anger, happiness, sadness, disgust, surprise, fear and neutral. Quantitative analyses are conducted to evaluate the performance of the proposed model as well as the quality of the yielded facial expression images. As evidenced by the empirical results, the proposed modified DCGAN model shows promising performance for realistic synthetic facial expression image generation.