Abstract

This research delves into the contemporary landscape of emotion recognition models trained on the FER 2013 dataset. The project introduces a real-time emotion detection system coupled with personalized song recommendations from Spotify. The model architecture, leveraging Keras and TensorFlow, is outlined, along with the challenges encountered in deployment. The dataset's characteristics, imbalances, and the model's training process are detailed. Furthermore, the paper discusses the incorporation of database functionality for improved song recommendations, scheduled playlist updates, and the aspiration to shift video streaming to the client side for deployability. Acknowledging the current accuracy of approximately 66%, the study proposes further training and potential exploration of alternative models, such as the Vision Transformer Model.