This paper is a comparative study on facial expression classification for emotion recognition using Convolutional Neural Networks (CNNs). The study evaluates both custom-designed CNN architectures and pre-trained models, such as VGG-16, MobileNetV2, and ResNet15, on a dataset of 35,000 grayscale images in seven emotion classes. The research focuses on the optimization of model performance through hyperparameter tuning, data augmentation, and preprocessing techniques. Key contributions include a in-depth analysis of feature extraction, transfer learning efficiency, and model generalization. A comprehensive evaluation using accuracy, precision, recall, F1-Score, and statistical significance tests is performed to validate performance differences. Results show that despite the Custom-CNN having competitive accuracy, pre-trained models such as VGG-16 outperform generalization. The study further gives insights in misclassification patterns through confusion matrix analysis, which presents different challenges in distinguishing subtle emotions like fear and surprise. Potential future works are in exploring Vision Transformers (ViTs) and hybrid models to further boost the facial expression recognition performance.