The study includes a complete design process for a facial expression recognition system that uses deep learning with image preprocessing algorithms for evaluation. The system uses grayscale conversion in combination with histogram equalization and face detection preprocessing that leads to feature extraction through convolutional neural networks (CNNs). The system underwent training and testing across two popular datasets named FER-2013 and CK+ because they offer diverse contents and high-quality annotations. This research utilized accuracy, precision, recall and F1-score for evaluating the system. The FER-2013 dataset evaluation produced 93.4% model accuracy together with 92.1% precision and 91.6% recall and 91.8% F1-score. The CK+ dataset allowed the model to achieve 96.2% accuracy and 95.7% F1-score thus demonstrating effective data generalization capabilities. The training stability increased along with overfitting reduction through the implementation of dropout together with batch normalization layers. © 2025 IEEE.