Development of a Robust Multi-Scale Featured Local Binary Pattern for Improved Facial Expression Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial expression recognition (FER) is crucial for AI and robotics.
• Traditional Local Binary Patterns (LBP) in FER struggle with scale-dependent texture information due to pixel loss.
• Existing LBP methods face limitations in capturing comprehensive facial texture features.

Purpose of the Study:

• To develop an improved LBP method for enhanced facial expression recognition.
• To address the limitations of traditional LBP in handling multi-scale texture information in facial images.
• To propose a novel feature extraction technique for more accurate FER.

Main Methods:

• A new extended LBP method, multi-scale featured local binary pattern (MSFLBP), was developed using bitwise AND operations on LBP(8,1) and LBP(8,2).
• Facial components (eyes, nose, lips) were detected, cropped, and sharpened using an unsharp masking kernel.
• The proposed MSFLBP method was evaluated using four machine learning classifiers on the CK+ and KDEF datasets.

Main Results:

• The MSFLBP method, particularly with Support Vector Machine (SVM), demonstrated superior performance compared to existing LBP-based approaches.
• Achieved 99.12% accuracy on the Extended Cohn-Kanade (CK+) dataset.
• Attained 89.08% accuracy on the Karolinska Directed Emotional Faces (KDEF) dataset.

Conclusions:

• The proposed MSFLBP method effectively overcomes the limitations of traditional LBP in facial expression recognition.
• MSFLBP combined with SVM offers a highly accurate and robust solution for FER.
• This approach significantly advances the state-of-the-art in facial expression analysis.