A study on computer vision for facial emotion recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial emotion recognition (FER) is a key area within computer vision.
• Deep neural networks (DNNs) show promise for FER tasks.
• Understanding critical facial features used by DNNs is essential for model interpretability and improvement.

Purpose of the Study:

• To identify critical facial features utilized by a deep neural network (DNN) for facial emotion recognition (FER).
• To evaluate the performance of a combined convolutional neural network (CNN) model on FER tasks.
• To assess the impact of cross-database validation on model accuracy.

Main Methods:

• A deep neural network (DNN) model, specifically a combination of squeeze-and-excitation network and residual neural network, was employed for FER.
• The model was trained and validated using two large-scale facial expression databases: AffectNet and the Real-World Affective Faces Database (RAF-DB).
• Feature maps from residual blocks were analyzed to pinpoint critical facial landmarks.

Main Results:

• Analysis indicated that facial features around the nose and mouth are critical landmarks for the DNN model in FER.
• A network model trained on AffectNet achieved 77.37% accuracy when validated on RAF-DB.
• Transfer learning on RAF-DB after pretraining on AffectNet yielded a higher validation accuracy of 83.37%.

Conclusions:

• The nose and mouth regions are identified as key facial areas for accurate FER by DNNs.
• Cross-database validation demonstrates the generalizability and robustness of the developed FER model.
• This research contributes to a better understanding of DNNs in computer vision and offers insights for enhancing FER system performance.