MobileNetV3-Based Fine-Tuned Facial Emotion Recognition with Targeted Class Balancing

Title

MobileNetV3-Based Fine-Tuned Facial Emotion Recognition with Targeted Class Balancing

Creator

Chaudhary, Anubhav; Poonia, Ramesh Chandra; Shanbhog, Manjula

Description

Facial emotion recognition (FER) is a pillar of affective computing and augmented human computer interaction, but has been stymied by the problem of class imbalance and lack of prevalence of subtle emotional differences. This paper presents a lightweight FER framework based on the MobileNetV3 architecture with a fine-tuned and weighted dataset that applies class balance and class weighting as strategies that optimized the three-class classification of three discrete emotions Angry, Happy, and Sad. The characteristics of the dataset were assembled comprising a total of 7,305 labelled facial images, based on the KDEF, Kaggle, and Face Expression Dataset hence inheriting the heterogeneity of subjects and imaging conditions. The pre-processing of all of the images carried out as the RGB input and after resizing (224 x 224 pixels) a massive data augmentation done to encourage generalization. Transfer learning in the training pipeline is done through progressive unfreezing and the weight of the loss on the minority classes (Angry and Sad) are boosted to improve the performance of detection. The achieved model resulted in an accuracy of 87% on the test set, and had equal accuracy in preciseness, recall, and F1-scores over all emotion types. Extended error analysis revealed that the majority of cases that were misclassified fell between the categories Angry and Sad because they were mistaken due to combining visual cues. Even then, the performance showed stability despite the variable lighting as well as in variable positional context. In Comparison, MobileNetV3 outperforms state-of-art-lightweight models with respect to accuracy and computation of similar computational complexity. 2025 IEEE.

Source

Proceedings of the 2025 International Conference on Artificial Intelligence and Emerging Technology, Global AI Summit 2025;pp.1574-1579

Date

01-01-2025

Publisher

Institute of Electrical and Electronics Engineers Inc.

Subject

Convolutional Neural Network; Facial Expression Recognition; MobileNetV3; Transfer Learning; VGG16

Coverage

Chaudhary A., Christ University, Department of Computer Science, Bangalore, India; Poonia R.C., Christ University, Department of Computer Science, Bangalore, India; Shanbhog M., Christ University, Department of Computer Science, Bangalore, India

Rights

Restricted Access; Hardcopy may be available in the library

Relation

ISBN: 979-833155379-1;

Format

online

Language

English

Type

Conference paper

Identifier

https://doi.org/10.1109/AISummit66170.2025.11411128

https://www.scopus.com/pages/publications/105035308851?origin=resultslist