This paper presents an enhanced real-time facial expression recognition (FER) system using deep learning techniques, specifically focusing on the extraction and categorization of seven universal emotional states: happiness, sadness, anger, fear, surprise, disgust, and neutrality. The study aims to improve the accuracy of FER in real-world datasets, which are more complex and less controlled compared to laboratory settings. The proposed method employs convolutional neural networks (CNNs), particularly MobileNet, to process images captured in real-time, addressing challenges such as degraded image quality, occlusions, variable lighting, and head pose variations. The effectiveness of the deep learning-based approach is rigorously evaluated through a comparative analysis with manual techniques and existing approaches, demonstrating superior accuracy and efficiency. The results show that the proposed model achieves an accuracy of 97.9% during training and 97.7% during validation, outperforming other models in the literature. The study highlights the potential of deep learning in enhancing FER capabilities, making it valuable for applications in human-computer interaction, market research, and mental health assessments.This paper presents an enhanced real-time facial expression recognition (FER) system using deep learning techniques, specifically focusing on the extraction and categorization of seven universal emotional states: happiness, sadness, anger, fear, surprise, disgust, and neutrality. The study aims to improve the accuracy of FER in real-world datasets, which are more complex and less controlled compared to laboratory settings. The proposed method employs convolutional neural networks (CNNs), particularly MobileNet, to process images captured in real-time, addressing challenges such as degraded image quality, occlusions, variable lighting, and head pose variations. The effectiveness of the deep learning-based approach is rigorously evaluated through a comparative analysis with manual techniques and existing approaches, demonstrating superior accuracy and efficiency. The results show that the proposed model achieves an accuracy of 97.9% during training and 97.7% during validation, outperforming other models in the literature. The study highlights the potential of deep learning in enhancing FER capabilities, making it valuable for applications in human-computer interaction, market research, and mental health assessments.