Image-based facial emotion recognition using convolutional neural network on emognition dataset

Area of Science:

• Computer Science
• Artificial Intelligence
• Cognitive Science

Background:

• Facial emotion recognition (FER) is challenging due to expression variability.
• Existing deep learning FER models are limited by restricted datasets.
• The Emognition dataset offers a broader range of ten target emotions.

Purpose of the Study:

• To expand deep learning applications for FER.
• To develop robust Convolutional Neural Network (CNN) models for classifying ten emotions.
• To evaluate model performance using transfer learning and custom architectures.

Main Methods:

• Data preprocessing: video to image conversion and data augmentation.
• Model development: transfer learning (Inception-V3, MobileNet-V2) and custom CNNs.
• Hyperparameter optimization using the Taguchi method for robust model building.

Main Results:

• Achieved high performance on the test dataset.
• Demonstrated an accuracy of 96%.
• Obtained an average F1-score of 0.95.

Conclusions:

• The proposed CNN models show significant potential for accurate FER.
• The study validates the effectiveness of deep learning on diverse emotion datasets.
• This research contributes to advancing the field of automated emotion detection.