Hierarchical Attention-Based Age Estimation and Bias Analysis.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Accurate age estimation from facial images is crucial for various applications.
• Existing methods face challenges in handling image variations and achieving high precision.

Purpose of the Study:

• To develop a novel Deep Learning approach for precise age estimation from facial images.
• To introduce an attention-based image augmentation-aggregation technique and a hierarchical probabilistic regression model.

Main Methods:

• Utilized a Transformer-Encoder for adaptive aggregation of multiple image augmentations.
• Developed a hierarchical probabilistic regression model combining discrete estimates and an ensemble of regressors.
• Trained regressors to refine probability estimates within specific age ranges.

Main Results:

• The proposed age estimation scheme significantly outperforms current state-of-the-art methods.
• Achieved new state-of-the-art accuracy on the MORPH II and CACD datasets.
• Presented an analysis of biases present in current age estimation techniques.

Conclusions:

• The novel Deep Learning approach offers superior accuracy for facial age estimation.
• The attention-based augmentation and hierarchical regression model are effective for this task.
• The findings contribute to advancing the field of automated facial age analysis.