[论文解读] Finger-GAN: Generating Realistic Fingerprint Images Using Connectivity Imposed GAN

Shervin Minaee, AmirAli Abdolrashidi|arXiv (Cornell University)|Dec 25, 2018

Biometric Identification and Security参考文献 34被引用 47

一句话总结

Finger-GAN 使用 DC-GAN 并结合一个总变差正则化项来生成真实且连通的指纹图像，在 FVC-2006 和 PolyU 数据集上进行评估，获得具有竞争力的 FID 分数。

ABSTRACT

Generating realistic biometric images has been an interesting and, at the same time, challenging problem. Classical statistical models fail to generate realistic-looking fingerprint images, as they are not powerful enough to capture the complicated texture representation in fingerprint images. In this work, we present a machine learning framework based on generative adversarial networks (GAN), which is able to generate fingerprint images sampled from a prior distribution (learned from a set of training images). We also add a suitable regularization term to the loss function, to impose the connectivity of generated fingerprint images. This is highly desirable for fingerprints, as the lines in each finger are usually connected. We apply this framework to two popular fingerprint databases, and generate images which look very realistic, and similar to the samples in those databases. Through experimental results, we show that the generated fingerprint images have a good diversity, and are able to capture different parts of the prior distribution. We also evaluate the Frechet Inception distance (FID) of our proposed model, and show that our model is able to achieve good quantitative performance in terms of this score.

研究动机与目标

• Motivate synthetic fingerprint generation to aid biometric research and testing.
• Develop a GAN-based framework that produces realistic fingerprint images resembling real samples.
• Impose connectivity in generated fingerprints via a regularization term to encourage connected ridge lines.
• Evaluate image realism and diversity using quantitative metrics and visual analysis on public fingerprint databases.

提出的方法

• Use a deep convolutional GAN (DC-GAN) as the core generator and discriminator.
• Introduce a regularization term based on total variation (TV) to promote connectivity of fingerprint ridges in the generated images.
• Train the model end-to-end with standard GAN objectives augmented by the TV regularizer: L_GAN-TV = E[log D(x)] + E[log(1−D(G(z)))] + λ TV(G(z)).
• Adopt a 4-layer discriminator and a 5-layer generator architecture with batch normalization and Leaky ReLU activations.
• Train on 64x64 central crops of two fingerprint databases (FVC-2006 DB2-A and PolyU DB-II) with 120 epochs, batch size 40, and Adam optimizer.

实验结果

研究问题

• RQ1 Can a GAN learn to generate fingerprint-like textures that are visually similar to real fingerprints?
• RQ2 Does adding a total variation-based connectivity regularizer improve ridge connectivity and realism in generated fingerprints?
• RQ3 How diverse and realistic are the synthetic fingerprints across different latent inputs and training epochs?
• RQ4 What is the quantitative performance of Finger-GAN as measured by Fréchet Inception Distance (FID) on real-like fingerprint samples?

主要发现

模型/数据库	Frechet Inception Distance
DC-GAN/ FVC Fingerprint	70.5

• The Finger-GAN generates fingerprint images that look very realistic and resemble real samples from the tested databases.
• Generated images show noticeable diversity across samples and improve in realism over training epochs.
• The TV regularization promotes connected ridge lines, aiding the natural connectivity seen in fingerprints.
• FID scores indicate competitive quantitative performance relative to state-of-the-art GAN models on public image datasets.
• Discriminator and generator losses steadily evolve during training, reflecting learning dynamics on both datasets.
• Visual results across epochs demonstrate gradual sharpening and realism of fingerprint details.