Multi-Task Convolutional Neural Network for Pose-Invariant Face Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Face recognition systems struggle with variations in pose, illumination, and expression (PIE).
• Multi-task learning (MTL) offers a promising avenue for improving model robustness by learning related tasks simultaneously.
• Existing MTL methods face challenges in balancing the contributions of different tasks.

Purpose of the Study:

• To develop an effective multi-task learning framework for robust face recognition.
• To address the challenge of balancing diverse tasks within a single network.
• To investigate the role of auxiliary tasks in disentangling identity features from PIE variations.

Main Methods:

• Proposed a multi-task convolutional neural network (CNN) integrating identity classification with PIE estimation.
• Developed a dynamic-weighting scheme for adaptive task weight assignment in MTL.
• Introduced a pose-directed CNN for learning pose-specific identity features within a joint framework.
• Utilized an energy-based weight analysis to understand the internal workings of the MTL CNN.

Main Results:

• The proposed dynamic-weighting scheme effectively balanced multiple tasks in the CNN.
• Pose-directed learning enabled the extraction of pose-invariant identity features.
• Auxiliary PIE tasks acted as regularizers, improving identity feature disentanglement.
• Achieved state-of-the-art or comparable performance on benchmark datasets (multi-PIE, LFW, CFP, IJB-A).

Conclusions:

• The proposed MTL approach significantly enhances face recognition accuracy, particularly under challenging PIE variations.
• The dynamic-weighting and pose-directed strategies are crucial for effective task integration.
• This work provides a novel and effective method for pose-invariant face recognition using comprehensive multi-PIE data.