Pix2pix network for fingerprint texture image synthesis

GANs (Generative Adversarial Networks) were proposed to generate realistic synthetic images. In this work, we will discuss the use of GANs as alternative reconstruction of different fingerprint images from the original ones. The samples result in the same person fingerprint but obtained with other textures. Thus, it is intended to contribute to improving the method to increase databases with new samples, incorporating textures, when the quantities are insufficient for any purpose. To verify the similarity of the synthesized images with the original ones, a convolutional Xception network and the RMSE metric are used. The results obtained with fingerprint images of 3 persons, 20 of each finger, and 4 different textures, showed the tradeoff between similarity, recognizability, and the number of epochs of the Pix2pix training.