PrintsGAN: Synthetic Fingerprint Generator.

Area of Science:

• Biometrics
• Computer Vision
• Machine Learning

Background:

• Limited availability of large-scale, diverse fingerprint datasets hinders research in fingerprint recognition.
• Existing datasets lack sufficient identities and impressions per finger, restricting studies on deep learning embeddings.

Purpose of the Study:

• To introduce PrintsGAN, a synthetic fingerprint generator, for creating large-scale, realistic fingerprint datasets.
• To demonstrate the utility of the generated synthetic dataset in training deep learning models for fingerprint recognition.

Main Methods:

• Development of PrintsGAN, a novel synthetic fingerprint generation method.
• Synthesis of a large dataset comprising 525,000 fingerprints (35,000 distinct fingers, 15 impressions each).
• Training and fine-tuning a deep network embedding model using the synthetic and real fingerprint data.

Main Results:

• The synthetic dataset enabled training of a deep network achieving 87.03% True Accept Rate (TAR) at 0.01% False Accept Rate (FAR) on the NIST SD4 database.
• This performance represents a significant improvement over models trained solely on existing real-world datasets (73.37% TAR).
• PrintsGAN's ability to generate multiple impressions per finger is crucial for achieving these performance gains.

Conclusions:

• PrintsGAN effectively addresses the scarcity of large-scale fingerprint datasets.
• The synthetic dataset generated by PrintsGAN significantly enhances the performance of deep learning-based fingerprint recognition systems.
• The publicly released dataset and generation method will advance research in biometrics and pattern recognition.