Facial Expression Recognition (FER) is a challenging task with the goal of classifying an individual’s emotional state from facial expressions. In this research work, the introduced method offers a robust solution to FER with the FER2013 dataset, spanning a broad spectrum of facial expressions in grayscale faces. The model is interested in the classification of seven main emotions: anger, disgust, fear, happiness, sadness, surprise, and neutrality. State of the art techniques are applied, such as the application of Optuna for hyperparameter optimization, customized loss functions to deal with class imbalance, and sophisticated data augmentation strategies. Such techniques collectively advance the scalability and generalizability of the model under diverse conditions. Leverage a deep neural network architecture together with transformers, the introduced approach achieves a test accuracy of 89.5%, outperforming conventional baseline methods. Important hyperparameters such as filter sizes, dense units, dropout rates, and learning rates—are optimized to achieve maximum feature extraction while preventing overfitting. The incorporation of class weights also mitigates the impacts of imbalanced datasets, encouraging recognition performance. Applications of this research work are in mental health screening, personalized learning devices, and human-computer interaction. Future work is in the extension of the framework to dynamic emotion recognition and cross domain applications, a significant leap in adaptive and efficient FER systems.