Improving Face-Based Age Estimation With Attention-Based Dynamic Patch Fusion.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are popular for face-based age estimation.
• Current CNN methods treat all facial regions equally, missing age-specific information.
• A need exists for methods that focus on important facial patches for age estimation.

Purpose of the Study:

• To propose an Attention-based Dynamic Patch Fusion (ADPF) framework for face-based age estimation.
• To improve age estimation accuracy by dynamically identifying and prioritizing age-specific facial patches.
• To overcome the limitations of existing methods that ignore the varying importance of facial regions.

Main Methods:

• Developed a novel framework, Attention-based Dynamic Patch Fusion (ADPF), comprising two CNNs: AttentionNet and FusionNet.
• Introduced a Ranking-guided Multi-Head Hybrid Attention (RMHHA) mechanism within AttentionNet to locate and rank age-specific patches.
• Implemented a diversity loss to encourage the discovery of diverse and important facial patches with reduced overlap.

Main Results:

• The proposed RMHHA mechanism dynamically ranks facial patches by importance.
• The FusionNet utilizes these ranked patches and the facial image for age prediction, with learning path length proportional to patch information content.
• Extensive experiments demonstrated that ADPF surpasses state-of-the-art methods on multiple age estimation benchmark datasets.

Conclusions:

• The ADPF framework effectively addresses the limitations of equal region treatment in existing CNN-based age estimation methods.
• The novel attention and fusion mechanisms significantly enhance the accuracy of face-based age estimation.
• ADPF represents a significant advancement in age estimation technology by intelligently leveraging facial patch information.