Evaluating MobileNetV2 Architecture for Resource-Efficient Facial Emotion Recognition

Abstract

Facial Emotion Recognition (FER) is a vital field within computer vision, focusing on classifying emotions based on facial expressions. This study investigates the performance of the MobileNetV2 model for FER, leveraging transfer learning to enhance its effectiveness. Pre-trained on the ImageNet dataset, MobileNetV2 was fine-tuned with additional layers to classify seven emotional categories from the FER-2013 dataset: happiness, sadness, anger, fear, surprise, disgust, and neutrality. Performance was evaluated using key metrics, including accuracy, AUC, and RMSE. The model achieved an accuracy of 73.82%, an AUC of 95.16%, and an RMSE of 11.37. These results underscore its efficiency and suitability for devices with limited computational resources, making it an ideal candidate for real-time applications. This research highlights MobileNetV2's potential as a cost-effective and efficient solution for emotion analysis in human–computer interaction.