Comparative Performance Analysis of Xception and ResNet50 Architectures for Facial Expression Recognition Using the FER-2013 Dataset

Abstract

This study aims to evaluate and compare the performance and computational efficiency of the Xception and ResNet50 architectures in facial expression classification tasks. Facial Expression Recognition (FER) plays an important role in the development of intelligent systems capable of interpreting human emotions. This technology has various applications, including adaptive learning systems, emotion-aware customer service, and mental health support systems. This research compares two widely used Convolutional Neural Network (CNN) architectures, Xception and ResNet50, for facial expression classification using the FER-2013 dataset. The dataset contains 35,887 grayscale facial images with a resolution of 48×48 pixels categorized into seven basic emotions. All images were resized to 224×224 pixels and converted into RGB format to match the input requirements of pretrained ImageNet models.

Both architectures were trained using a transfer learning strategy with selective fine-tuning on specific layers. Data augmentation techniques were applied to increase dataset variability and reduce overfitting. Model performance was evaluated using accuracy, precision, recall, F1-score, and confusion matrix metrics. The results show that Xception architecture outperforms ResNet50, achieving a validation accuracy of 70.69% and a weighted F1-score of 0.71. These findings demonstrate that appropriate architecture selection and structured training strategies can significantly improve FER performance in practical intelligent systems.