Multiscale wavelet attention convolutional network for facial expression recognition

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial expression recognition (FER) is crucial for human-computer interaction.
• Current Convolutional Neural Networks (CNNs) require accuracy enhancements for robust FER applications.

Purpose of the Study:

• To improve the accuracy of facial expression recognition systems.
• To introduce novel deep learning architectures for enhanced FER.

Main Methods:

• Proposed Multi-scale CNN (MCNN) by replacing the first convolutional layer with a Multi-scale Convolutional (MsC) layer.
• Introduced wavelet Channel Attention CNN (wCA-CNN) by incorporating a wavelet Channel Attention (wCA) mechanism.
• Developed wCA-based Multi-scale CNN (wCA-MCNN) combining MsC and wCA.
• Applied these methods to baseline Residual Network (ResNet18).

Main Results:

• MCNN improved accuracy by 1.339% over CNN.
• wCA-CNN improved accuracy by 1.414% over CNN.
• wCA-MCNN achieved a 2.921% accuracy improvement over CNN.
• ResNet18 variants showed improvements up to 1.810%.

Conclusions:

• The proposed MCNN, wCA-CNN, and wCA-MCNN architectures significantly enhance facial expression recognition accuracy.
• The integration of MsC layers and wCA mechanisms offers a promising direction for advancing FER systems.
• The methods were validated on both real-world (FESR) and standard (KDEF) datasets.