Uma abordagem para geração de imagens baseada no uso de GPU e redes neurais artificiais

Abstract

With the crescent demand for embedded computer vision solutions, a strategic functional verification is increasingly needed. In this context,the present work aims at thespecification and implementation of a synthetic image generator that produces images derived from initial image datasets. This work includes a bibliographical research in the works of the pertinent scienti?c literature, details of an implementation and also an experimental evaluation to present more information about the present proposal. The process of generation of the derived images was conceived through components that work with methods of generationby deformation and generation by artificial neural networks. The developed components were designed with parallel computing, using the CUDA platform, as well as using TensorFlow for implementations of the neural networks involved. There were implemented Convolutional Neural Network (CNN) and Generative Adversarial Networks (GAN) in one of the methods of image generation. The results about the implemented component corroborate the feasibility of its use in the field of data augmentation, in functional distributed verifications and in the training of artificial neural networks.