Face Attribute Estimation Using Multi-Task Convolutional Neural Network.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Face attribute estimation is crucial for applications like face recognition and marketing analysis.
• Current Convolutional Neural Networks (CNNs) often share feature extractors for efficiency, potentially limiting accuracy.
• Optimizing CNN structures for multiple binary classifications is key to improving face attribute estimation.

Purpose of the Study:

• To propose a novel method, Merged Multi-CNN (MM-CNN), for optimizing CNN structures in face attribute estimation.
• To enhance both parameter efficiency and estimation accuracy.
• To introduce Convolutionalization for Parameter Reduction (CPR) to further optimize MM-CNNs.

Main Methods:

• Developed Merged Multi-CNN (MM-CNN) for automated CNN structure optimization.
• Implemented Convolutionalization for Parameter Reduction (CPR) to remove fully connected layers.
• Conducted experiments on CelebA and LFW-a datasets to evaluate performance.

Main Results:

• MM-CNN with CPR demonstrated superior efficiency in face attribute estimation.
• The proposed method achieved higher estimation accuracy compared to conventional methods.
• Significant reduction in the number of weight parameters was observed.

Conclusions:

• MM-CNN combined with CPR offers an effective solution for accurate and parameter-efficient face attribute estimation.
• The approach addresses limitations of shared feature extractors in traditional CNN-based methods.
• This work advances the field of facial analysis through optimized deep learning architectures.