Facial Expression Recognition Using Local Sliding Window Attention.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial expression recognition (FER) faces significant challenges due to facial occlusion and head pose variations, which compromise facial information and complicate feature extraction.
• Existing FER methods often rely on prior knowledge or fixed-size local cropping, leading to either complex preprocessing or destruction of local feature integrity.

Purpose of the Study:

• To introduce a novel Sliding Window Attention Network (SWA-Net) designed to overcome the limitations of current facial expression recognition techniques.
• To enhance the extraction of fine-grained facial features despite challenges like occlusion and head pose variations.

Main Methods:

• Proposed a Sliding Window Attention Network (SWA-Net) employing a sliding window strategy for feature-level cropping, preserving local feature integrity without complex preprocessing.
• Integrated a local feature enhancement module utilizing a multiscale deep network to mine fine-grained features with intraclass semantics.
• Introduced an adaptive local feature selection module to guide the model in identifying essential local features.

Main Results:

• The SWA-Net model achieved high accuracy on benchmark datasets, demonstrating comparable performance to state-of-the-art methods.
• Achieved scores of 90.03% on RAF-DB, 89.22% on FERPlus, and 63.97% on AffectNet, validating the model's effectiveness.

Conclusions:

• The proposed SWA-Net effectively addresses key challenges in facial expression recognition, particularly occlusion and head pose variations.
• The sliding window attention mechanism and feature enhancement modules contribute to robust and accurate facial expression recognition.