Comparative Analysis of Deep Neural Networks for Online Facial Expression Recognition and Facial Shape Classification

Abstract

In the context of affective computing, real-time facial analysis plays an important role in human–computer interaction and personalized applications. This study presents a controlled comparative evaluation of five widely used deep neural network architectures EfficientNetB0, VGG16, MobileNetV2, InceptionV3, and ResNet50 for two classification tasks: recognizing eight facial-expression categories and seven face-shape categories using the mtr_face-v6 dataset. An online recognition pipeline was developed to investigate the trade-off between classification performance and computational efficiency. The experimental results demonstrated that InceptionV3 achieved the highest facial-expression recognition accuracy of 97.4%, benefiting from its inherent multi-scale feature-extraction capability. EfficientNetB0 achieved the highest face-shape classification accuracy of 97.0%, owing to its compound scaling mechanism and ability to learn global facial geometry. Although MobileNetV2 produced lower accuracies of 90.0% for facial-expression recognition and 88.4% for face-shape classification, it provided the lowest inference latency and remained suitable for resource-constrained online applications. Based on these findings, a task-specialized dual-model framework was adopted, integrating InceptionV3 for facial-expression recognition and EfficientNetB0 for face-shape classification within a unified graphical interface. Overall, the results demonstrate that task-specific model selection can achieve high classification performance for both facial-expression recognition and face-shape classification. Future work will investigate a unified multi-task learning architecture to jointly optimize both tasks within a shared framework. This study provides practical guidance for developers and researchers in selecting suitable deep-learning backbones according to the accuracy and computational requirements of online facial-analysis systems.