Age Estimation of Faces in Videos Using Head Pose Estimation and Convolutional Neural Networks

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Accurate age estimation from human faces is crucial but difficult due to variations in apparent vs. physical age.
• Current deep learning methods struggle with age estimation in videos due to changing head poses across frames.

Purpose of the Study:

• To enhance the performance of age estimation for faces in videos by integrating head pose information.
• To develop a robust system that accounts for head pose variations in facial age estimation.

Main Methods:

• A combined system utilizing deep regression forests (DRFs) for age estimation and a multiloss convolutional neural network (CNN) for head pose estimation was developed.
• Age estimation was refined by considering only faces within a specific head pose degree threshold.
• Datasets like CACD and AFAD were used, with images segmented by estimated head pose to analyze performance variations.

Main Results:

• Age estimation accuracy was significantly higher for frontal facial images compared to those at various angles, confirming the impact of head pose.
• The proposed combined system demonstrated superior precision and reliability in age estimation for videos compared to methods lacking head pose analysis.

Conclusions:

• Head pose is a critical factor influencing the accuracy of facial age estimation in videos.
• Integrating head pose estimation into age estimation systems offers a more accurate and reliable solution for video-based applications.