The work integrates a real-time facial emotion recognition system using the RAF-DB dataset, which contains 15,350 images annotated for seven basic emotions. Different aspects of dataset preparation and data augmentation, architectures to be used for the model, and performance evaluation on emotion classification are discussed within this study. A CNN is used in the emotion detection that can be drawn from live video streams in very high efficiency. Even though it is varied and contains images of various age ranges and demographic differences, the dataset nonetheless presents challenges like class imbalance and issues regarding the privacy and bias of data. Significant improvement in a model's generalization ability can be seen after using data augmentation techniques like rescaling, shearing, zooming, and horizontal flipping. Best accuracy was obtained at epoch 25, which was 78.32% for validation with three hidden layers and a filter of 3x3. This model therefore maintains an equilibrium between accuracy and real-time performance. The performance of the proposed model was tested across various demographic groups, and also with variations in the accuracy presented in most cases of the detection of fear and disgust. The research has compared its proposed model with existing models such as VGGNet and ResNet to give prominence to the computational efficiency, making it fit for real-time applications in situations where a dataset would be relatively small and limits the amount of computational resources available. The findings are important to put emotions recognition into practice in the real world, bringing to limelight the need for balancing model performance with ethical applicability in the field of AI.

Real-time Emotion Detection CNN Deep Learning Imbalanced Dataset Facial Expression Recognition Accuracy