The Face Generation for Emotion Analysis Using Deep Neural Networks

Facial expression recognition (FER) using deep learning from short training sets remains challenging despite its significance to human-machine interactions. This study provides a unique generative adversarial network approach for transforming expressionless face photographs into those depicting a range of emotions to address the problem of insufficient training data for deep learning for FER. A novel GAN architecture is devised, using a generator with two coding devices with two decoders, two discriminators, and a feature extractor, to combat the distortions and overfitting that may result from training GANs on small training datasets. In order to accomplish this, the proposed GAN architecture makes use of the mechanism of a feature map to discover regional include deviations among images in both the source and target domain, which not only allows it to generate appealing facial expression pictures, but also to preserve the initial signs in input neutral pictures of faces. The suggested GAN was utilised to add photographs with up to seven distinct face expressions to the training dataset, which improved the FER accuracy of the deep neural networks tested.