Patch Attention Layer of Embedding Handcrafted Features in CNN for Facial Expression Recognition

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Facial expression recognition is vital for human-computer interaction.
• Traditional methods struggle with shallow feature extraction and uniform weight sharing.
• Convolutional layers may not optimally capture nuanced facial characteristics.

Purpose of the Study:

• To propose a novel Patch Attention Layer (PAL) for improved facial expression recognition.
• To learn local shallow facial features effectively from image patches.
• To overcome limitations of traditional handcrafted features and uniform weight sharing.

Main Methods:

• Extracting Gabor surface features (GSF) using Gabor filters.
• Segmenting features into non-overlapping patches for local analysis.
• Employing a Patch Attention Layer (PAL) to learn weighted shallow features.
• Integrating these features into convolutional layers for high-level feature extraction.

Main Results:

• Achieved high accuracy on benchmark datasets: CK+ (98.93%), Oulu-CASIA (97.57%), JAFFE (93.38%), and RAF-DB (86.8%).
• Demonstrated competitive performance compared to state-of-the-art methods.
• Method operates directly on static images, simplifying preprocessing.

Conclusions:

• The proposed Patch Attention Layer (PAL) method effectively captures local facial features for expression recognition.
• This approach offers a simplified, yet powerful, alternative to existing methods.
• The technique shows significant potential for real-world human-computer interaction applications.