DHI-GAN: Improving Dental-Based Human Identification Using Generative Adversarial Networks

Area of Science:

• Forensic Science
• Biometrics
• Artificial Intelligence

Background:

• Dental-based human identification (DHI) faces challenges with small sample sizes.
• Accurate identification is crucial in forensic and biometric applications.

Purpose of the Study:

• To propose a novel semisupervised framework to address the small-sample problem in DHI.
• To enhance DHI performance using a "classifying while generating" paradigm with a generative adversarial network (GAN).

Main Methods:

• Developed DHI-GAN, a generative adversarial network incorporating an additional classifier for efficient training.
• Implemented an identity embedding-guided architecture and a parallel spatial and channel fusion attention block for feature learning.
• Utilized a combination of ArcFace and focal loss, with parameters to control generated samples during optimization.

Main Results:

• The DHI-GAN framework achieved a top-one accuracy rate of 92.5% on a real-world dataset.
• Outperformed existing baseline methods in dental-based human identification.
• Demonstrated the effectiveness of the GAN-based semisupervised strategy in reducing the need for extensive training data.

Conclusions:

• The proposed DHI-GAN framework offers a powerful solution for small-sample DHI.
• The semisupervised training strategy can be integrated with other classification models.
• This approach significantly advances the field of forensic biometrics.