Real-Time Facial Emotion Recognition Using Mini-Xception and EfficientNetB4

Abstract

Facial Emotion Recognition (FER) is an important field within computer vision and human–computer interaction that focuses on the automatic recognition of human emotional expressions through facial images. This study presents a comparative analysis of two Convolutional Neural Network (CNN) architectures, namely Mini-Xception and EfficientNetB4, for real-time facial emotion classification using the RAF-DB (Real-world Affective Faces Database) dataset. Mini-Xception was employed as a lightweight model with lower computational requirements, whereas EfficientNetB4 utilized a transfer learning approach to achieve superior classification performance. The RAF-DB dataset consists of seven primary emotion categories: angry, disgust, fear, happy, neutral, sad, and surprise. The preprocessing stage included facial image resizing, grayscale conversion for Mini-Xception, RGB normalization for EfficientNetB4, and the application of data augmentation techniques to improve model generalization capability. Experimental results demonstrated that Mini-Xception achieved a validation accuracy of 52.12%, while EfficientNetB4 attained a validation accuracy of 86.02%. In real-time implementation using a webcam and OpenCV, Mini-Xception exhibited advantages in inference speed, whereas EfficientNetB4 produced more stable and accurate emotion predictions. The findings indicate a trade-off between computational efficiency and classification performance. Therefore, EfficientNetB4 is more suitable for systems requiring high classification accuracy, while Mini-Xception is more appropriate for real-time applications operating under limited computational resources.