Expression-Invariant Age Estimation Using Structured Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Biometrics

Background:

• Automatic age estimation is crucial for various applications.
• Facial expressions can significantly impact the accuracy of age estimation algorithms.
• Existing methods often struggle to disentangle age and expression cues.

Purpose of the Study:

• To develop an expression-invariant automatic age estimation method.
• To investigate the influence of facial expressions on age estimation accuracy.
• To propose a novel graphical model for joint age and expression learning.

Main Methods:

• A new graphical model with a latent layer was introduced to learn the relationship between age and expression.
• This model captures facial changes related to aging and expressions.
• Experiments were conducted on FACES, Lifespan, and NEMO datasets.

Main Results:

• Joint learning of age and expression significantly improved age estimation performance.
• Age estimation error was reduced by 14.43% (FACES), 37.75% (Lifespan), and 9.30% (NEMO).
• The proposed model outperformed existing methods without prior expression knowledge and showed improvements when incorporating gender.

Conclusions:

• Jointly learning age and expression leads to more robust and accurate age estimation.
• The proposed graphical model effectively achieves expression-invariant age estimation.
• The model's flexibility allows for the integration of additional cues like gender, further enhancing performance.