[Paper Review] Multi-Expert Gender Classification on Age Group by Integrating Deep Neural Networks

Jun Beom Kho|arXiv (Cornell University)|Sep 6, 2018

Face recognition and analysis33 references4 citations

TL;DR

This paper proposes Multi-Expert Gender Classification on Age Group (MGA), an end-to-end multi-task learning framework that integrates Convolutional Appearance Network (CAN) for facial appearance and Deep Geometry Network (DGN) for facial shape features, with age group classification guiding expert specialization. MGA achieves state-of-the-art gender classification accuracy (99.11%) on MORPH-2 and 93.80% on Adience with only 2M parameters, significantly outperforming prior methods while reducing computational cost by up to 69x.

ABSTRACT

Generally, facial age variations affect gender classification accuracy significantly, because facial shape and skin texture change as they grow old. This requires re-examination on the gender classification system to consider facial age information. In this paper, we propose Multi-expert Gender Classification on Age Group (MGA), an end-to-end multi-task learning schemes of age estimation and gender classification. First, two types of deep neural networks are utilized; Convolutional Appearance Network (CAN) for facial appearance feature and Deep Geometry Network (DGN) for facial geometric feature. Then, CAN and DGN are integrated by the proposed model integration strategy and fine-tuned in order to improve age and gender classification accuracy. The facial images are categorized into one of three age groups (young, adult and elder group) based on their estimated age, and the system makes a gender prediction according to average fusion strategy of three gender classification experts, which are trained to fit gender characteristics of each age group. Rigorous experimental results conducted on the challenging databases suggest that the proposed MGA outperforms several state-of-art researches with smaller computational cost.

Motivation & Objective

• Address the negative impact of facial aging on gender classification accuracy in unconstrained environments.
• Improve gender classification performance by leveraging age group-specific expertise and facial geometric features.
• Develop a lightweight, efficient deep learning model that jointly estimates age and performs gender classification.
• Integrate appearance and geometric features through a novel model integration strategy without increasing model size.
• Demonstrate superior performance on benchmark datasets with minimal computational cost compared to prior state-of-the-art methods.

Proposed method

• Utilize a Convolutional Appearance Network (CAN) based on lightweight CNN to extract facial appearance features such as skin texture and color.
• Propose a Deep Geometry Network (DGN) that uses half-face inputs to extract robust facial shape features, especially effective under pose variations.
• Implement a model integration strategy that jointly fine-tunes CAN and DGN to enhance feature fusion and overall performance.
• Categorize input images into three age groups (young, adult, elder) using age estimation, then apply age-specific gender classification experts.
• Apply an average fusion strategy across three age-group-specific gender experts to produce the final gender prediction.
• Train the entire MGA system end-to-end using multi-task learning for joint age estimation and gender classification.

Experimental results

Research questions

• RQ1 Can age group-aware specialization of gender classification experts improve performance on aging-affected facial data?
• RQ2 How effective is the integration of deep facial geometry features (via DGN) with appearance features (via CAN) for gender classification?
• RQ3 Does the proposed model integration strategy enhance performance without increasing model size?
• RQ4 Can MGA achieve state-of-the-art accuracy with significantly reduced parameter count compared to existing methods?
• RQ5 How does the performance of MGA vary across diverse datasets like Adience and MORPH-2 under real-world conditions?

Key findings

• MGA achieves 99.11% gender classification accuracy on MORPH-2, outperforming prior methods with only 2M parameters—69 times smaller than methods like [28] and [32].
• On the Adience dataset, MGA achieves 93.80% gender accuracy and 79.55% exact age estimation accuracy, surpassing state-of-the-art methods despite a smaller model size.
• The integration of DGN with CAN improves performance, demonstrating that geometric features enhance gender classification, especially under aging effects.
• When combined with ResNet-50, MGA (ResNet-50) achieves 99.11% gender accuracy and 2.95 MAE on MORPH-2, showing compatibility and scalability with deeper architectures.
• The average fusion of age-group-specific experts reduces performance degradation in young and elder age groups, confirming the effectiveness of age-aware specialization.
• The proposed model integration strategy successfully fuses appearance and geometric features without increasing model size, maintaining computational efficiency.