Facial Expression Recognition Robust to Occlusion and to Intra-Similarity Problem Using Relevant Subsampling

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Facial Expression Recognition (FER) faces challenges with occluded images and variations between individuals.
• Existing methods struggle with intra-similarity, misclassifying similar expressions on different faces or vice versa.

Purpose of the Study:

• To develop a robust FER system addressing occlusion and intra-similarity problems.
• To improve the accuracy and reliability of FER in unconstrained environments.

Main Methods:

• Utilizing a Spatial Transformer Network (STN) integrated with an attention mechanism to focus on relevant facial regions.
• Implementing a triplet loss function to effectively handle intra-similarity variations.
• Combining STN with attention and triplet loss for enhanced feature learning.

Main Results:

• The proposed FER approach demonstrates robustness against image occlusion.
• Achieved a 2.09% higher accuracy on CK+ datasets and 0.48% higher on FER2013 compared to existing methods.
• Outperformed traditional deep neural networks and classical FER techniques.

Conclusions:

• The integration of STN, attention, and triplet loss significantly improves FER performance, especially in practical, occluded conditions.
• The developed method offers a more accurate and reliable solution for recognizing facial expressions from diverse datasets.