Facial Expression Recognition Based on Weighted-Cluster Loss and Deep Transfer Learning Using a Highly Imbalanced...

Area of Science:

• Computer Vision
• Pattern Recognition
• Machine Learning

Background:

• Facial expression recognition (FER) faces challenges due to real-world complexities and imbalanced datasets.
• Deep convolutional neural networks (CNNs) show potential for FER but struggle with performance degradation.
• Existing facial emotion datasets often lack sufficient data and exhibit inherent imbalances.

Purpose of the Study:

• To develop an improved deep CNN-based model for accurate facial expression recognition.
• To address the performance degradation of FER models in real-world scenarios.
• To overcome limitations posed by limited training data and imbalanced emotion datasets.

Main Methods:

• Applied deep transfer learning techniques using a CNN pre-trained on the VGGFace2 database.
• Proposed a novel weighted-cluster loss function for the fine-tuning phase.
• The weighted-cluster loss enhances intra-class compactness and inter-class separability while accounting for dataset imbalance.

Main Results:

• The proposed weighted-cluster loss function significantly improved FER performance on the AffectNet dataset.
• The method demonstrated superior results compared to baseline CNN models using weighted-softmax or center loss.
• Achieved robust recognition of eight basic facial emotions in unconstrained environments.

Conclusions:

• The novel weighted-cluster loss function is effective in improving deep CNN-based facial expression recognition.
• The approach successfully tackles challenges of data imbalance and feature discriminability in FER.
• This method offers a promising solution for real-world facial emotion analysis.