A Lightweight GAN Network for Large Scale Fingerprint Generation

Generating fingerprint images for biometric purposes is both necessary and challenging. In this study, we presented a fingerprint generation approach based on generative adversarial network. To ensure GAN training stability, we have introduced conditional loss doping that allows a continuous flow of gradients. Our study utilizes a careful combination of a residual network and spectral normalization to generate fingerprints. The proposed average residual connection shows more immunity against vanishing gradients than a simple residual connection. Spectral normalization allows our network to enjoy reduced variance in weight generation, which further stabilizes the training. Proposed scheme uses spectral bounding only in the input and the fully connected layers. Our network synthesized fingerprints up to 256 by 256 in size. We used the multi-scale structural similarity (MS-SSIM) metric for measuring the diversity of the generated samples. Our model has achieved 0.23 MS-SSIM scores for the generated fingerprints. The MS-SSIM score indicates that the proposed scheme is more likely to produce more diverse images and less likely to face mode collapse.