Mental Health Prediction Using Human Emotions with Swin Transformer

Facial Expression Recognition (FER) has become a key part of affective computing. Facial cues offer dependable clues about a person's emotions and frame of mind. Given the increase in mental health concerns all over the world, methods for keeping track in a way that doesn't disrupt daily life have become more vital, especially in the workplace, healthcare, and research. This work puts emphasis on using vision models, in particular the Swin Transformer, to help spot mental health states by looking at facial emotions. These models use datasets like FER2013 and CK+, which allow the system to spot basic emotions such as happiness, sadness, anger, fear, surprise, and a neutral state. This project's goal is to create emotion-recognition systems that can precisely spot early signs of student stress, depression, and other risks to mental health in real-time. Current FER methods, namely those making use of CNNs, typically struggle with slight changes in facial expressions, perform strangely in uncontrolled settings, and don't explain how certain parts of the face lead to how emotions are predicted. This project also mentions problems, such as datasets that are not balanced, limited ability to adapt, and a disconnect between recognizing emotions and scoring mental health in a way that is helpful. To fix this, the system uses Swin Transformer models improved with attention and interpretation tools like Grad-CAM. The design includes a way to score mental health by using average confidence and the number of images looked at. By mixing feature extraction based on transformers, organized data preprocessing, and evaluation scripts, the repository gives a framework for emotion-based mental health studies without going into how the ideas are put into action.