Facial Expression Recognition on the High Aggregation Subgraphs.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial Expression Recognition (FER) performance has improved with deep learning, but challenges remain due to nonlinear expression changes.
• Existing Convolutional Neural Networks (CNNs) for FER often overlook crucial inter-expression relationships, hindering recognition of confusable expressions.
• Graph Convolutional Networks (GCNs) capture relationships but suffer from low subgraph aggregation, including uncertain neighbors and increasing learning complexity.

Purpose of the Study:

• To propose a novel FER method that addresses the limitations of existing CNN and GCN approaches.
• To improve FER accuracy and efficiency by modeling complex expression relationships using high aggregation subgraphs (HASs).
• To leverage the strengths of both CNNs for feature extraction and GCNs for graph pattern modeling.

Main Methods:

• Formulated FER as a vertex prediction problem.
• Utilized vertex confidence to identify high-order neighbors for constructing HASs based on top embedding features.
• Employed GCN for reasoning on HASs to infer vertex classes, avoiding extensive overlapping subgraphs.

Main Results:

• The proposed method effectively captures underlying relationships between facial expressions on HASs.
• Achieved higher recognition accuracy and improved efficiency compared to state-of-the-art FER methods.
• Demonstrated superior performance on both in-the-lab and in-the-wild datasets.

Conclusions:

• The developed HAS-based FER method significantly enhances recognition accuracy by modeling inter-expression relationships.
• The approach offers a more efficient and effective solution for FER, particularly for confusable expressions.
• Highlighting the critical role of underlying expression relationships in advancing FER technology.