Facial emotion recognition has become a crucial application area in artificial intelligence, allowing machines to interpret and respond to human emotions based on facial expressions. This capability is particularly valuable in fields such as mental health support, educational technology, security systems, and human-computer interaction. In this project, we present a deep learning-based system designed to predict human emotions from facial images using the advanced EfficientNetB7 architecture. By leveraging the power of transfer learning from large-scale datasets like ImageNet, combined with modern regularization techniques such as dropout, L2 penalty, and batch normalization, the model achieves high classification accuracy and demonstrates strong generalization performance on a curated facial emotion dataset. The project encompasses a comprehensive pipeline that includes data preprocessing, model architecture customization, training procedures, and evaluation metrics. A structured dataset is prepared with clear separation of training, validation, and test sets, and image data is normalized and resized to meet model input requirements. The training process is carefully monitored through visualization of learning curves and further supported by a confusion matrix that provides insights into class-specific performance and misclassification trends. Experimental results show that EfficientNetB7 is highly effective in extracting subtle facial features critical to emotion recognition, outperforming simpler CNN architectures. These findings highlight the model's robustness and potential applicability in real-world scenarios. Looking forward, enhancements such as expanding the dataset, applying more aggressive data augmentation, fine-tuning model parameters, and deploying the system in real-time environments could further improve performance and usability.