Desarrollo de una aplicación web que genere rostros de personas que no existen en el mundo real

Abstract

With the increasing development of Generative Adversarial Network (GAN), image generation is an exciting challenge in the field of deep learning and artificial intelligence. Internationally, several works related to generative adversarial network models have been developed, but a significant gap persists in the lack of comparisons between different algorithms. The ability to generate non-existent images that closely resemble real-world images is of interest for many use cases. In the present research, we propose to develop a web application based on pre-existing neural networks using antagonistic generative technology to generate faces of people. In this study, 5,000 faces belonging to the FFHQ Face Dataset (2,500 female face) and (2,500 male face) were considered as the study population. The techniques that were employed are based on two models using the following algorithms: Generative Adversarial Network (GAN) and Deep Convolutional Generative Generative Adversarial Network (DCGAN). Statistical measures of the confusion matrix as a result of classification were used as performance metrics. The results allow concluding that the model (DCGAN) is the best classification method, due to its better prediction of the obtained values of true positives, and true negatives, in addition, the average accuracy of the model was 85.02%, higher than the other model. The developed application has successfully fulfilled the main objective of generating human faces from artificial intelligence algorithms