Dataset Enlargement with Generative Adversarial Neural Networks

Addressing dataset imbalance remains a critical challenge in human facial emotion classification. The predominant dataset, Facial Emotion Recognition 2013 (FER2013), exhibits significant class disparity among its 28,709 training and 3,589 test samples across seven emotions: anger, disgust, fear, happiness, sadness, surprise, and neutrality. The disproportionate representation, notably the predominance of ‘Happy’ (7,215 samples) and scarcity of ‘Disgust’ (436 samples), biases classification models towards overestimating happiness and underestimating disgust. We introduce the Transfer-Learning with Filters Generative Adversarial Network (TLF-GAN) approach to counteract this imbalance. TLF-GAN synthesizes artificial samples for underrepresented classes to balance the dataset. A base GAN model is initially trained using a comprehensive artificial face dataset. Subsequently, this model undergoes further training on underrepresented class samples. This process is supplemented by integrating binary classification models as filters within the TLF-GAN’s loss function, refining the generated outputs towards desired characteristics. This methodology augments the representation of scarce classes and establishes a feedback mechanism, enabling the continuous generation of quality samples for further training enhancements. The TLF-GAN approach offers a promising solution to the dataset imbalance issue, paving the way for more equitable and accurate emotion classification models. Moreover, its implications extend beyond facial emotion recognition, offering a viable strategy for mitigating data scarcity across various domains within artificial intelligence. This research underscores the importance of balanced datasets in training AI models, potentially leading to more accurate and fair outcomes in automated emotion recognition and beyond.