Cost-Sensitive Local Binary Feature Learning for Facial Age Estimation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Facial age estimation is crucial for various applications.
• Traditional methods rely on hand-crafted or holistic features, which can be sensitive to variations.
• Existing approaches may not adequately address the cost-sensitive nature of age estimation.

Purpose of the Study:

• To propose a novel cost-sensitive local binary feature learning (CS-LBFL) method for facial age estimation.
• To learn discriminative local features directly from raw pixels, enhancing face representation.
• To improve robustness against illumination and expression variations compared to conventional methods.

Main Methods:

• Developed a CS-LBFL method using hashing functions to project raw pixel values into low-dimensional binary codes.
• Ensured similar chronological ages map to close binary codes and dissimilar ages to distant codes.
• Proposed a cost-sensitive local binary multi-feature learning approach using multi-scale face patches for complementary information.
• Pooled and encoded local binary codes into real-valued histogram features for final face representation.

Main Results:

• Achieved competitive performance on four widely used face aging datasets.
• Demonstrated the effectiveness of learning local binary features directly from pixels.
• Showcased the robustness of the proposed method to illumination and expression variations.

Conclusions:

• The CS-LBFL method offers a promising approach for accurate and robust facial age estimation.
• Learning local binary features directly from pixels is effective for face representation.
• The multi-feature learning strategy enhances performance by exploiting complementary information from different scales.