Robust RGB-D Face Recognition Using Attribute-Aware Loss.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Current convolutional neural network (CNN) face recognition methods optimize feature separation but can be biased by training data sampling and may not capture attribute relationships.
• Existing models trained solely on facial images may fail to correlate features with important attributes like gender and ethnicity, limiting application scope.

Purpose of the Study:

• To develop a novel CNN-based face recognition approach that integrates facial attributes into the training process.
• To enhance feature discriminability and correlation with attributes by employing an attribute-aware loss function.

Main Methods:

• Proposed a new CNN architecture incorporating an attribute-aware loss function to regularize feature mapping based on attribute proximity.
• Trained the face recognition model on a large-scale RGB-D dataset comprising over 100,000 identities captured under real-world conditions.
• Evaluated the approach against existing methods using diverse experimental settings.

Main Results:

• The proposed attribute-aware loss function significantly improves the discriminative power of learned features.
• Integration of the depth channel from RGB-D data enhances recognition accuracy and robustness.
• Experimental comparisons validate the superiority of the new approach over conventional methods.

Conclusions:

• Incorporating attribute information and depth data into CNN training leads to more accurate and robust face recognition.
• The attribute-aware loss function is effective in learning features that are better correlated with demographic and other facial attributes.