Synthetic Finger Print Image Generation Using Modified Deep Convolutional GAN

Fingerprint images are often used as an important biometric tool for authentication and verification of an individual. In finger print researches such as recognition problems, large realistic finger print dataset is required for testing and validation of new algorithms. But due to legal issues most biometric dataset is not publicly available. Synthetic finger print generation techniques can solve this issue. But most models lack resolution due to various reasons like small training sample and less efficient algorithms. In this paper a Generative Adversarial Network (GAN) based finger print generation is presented. A Deep Convolutional GAN(DC-GAN) with a modified loss function which takes both BCE loss and Hinge Embedding Loss for training is implemented. The generated fingerprint looks similar to human finger prints and shows better resolution in terms of ridge structure, ridge endings and bifurcations. Compared to GAN architecture which considers only BCE loss, this network has better convergence in terms of generator and discriminator loss. The quality of the generated images is analyzed in terms of structural similarity index and frechet inception distance.