Estudo sobre diferentes métodos de transferência de estilo com Redes Convolucionais e desenvolvimento de uma ferramenta para mostra acadêmica

Abstract

Humans have long mastered the ability to create visually unique works of art. Such a task is not easily reproducible by computers, especially when considering the concept of creating something original. On the other hand, algorithms have proven increasingly efficient at learning when trained on existing content databases. From this, studies on Neural Style Transfer (NST) emerged, consisting of combining the style of one image with the content of another, with the main purpose of generating new artistic images from existing artworks. The proposal of this work was to reproduce and compare the seminal algorithm by (GATYS et al., 2015) in different simulation scenarios. Comparative tests were performed to evaluate: (i) the influence of the initialization image (white noise versus content image); (ii) the performance of different optimizers (LBFGS versus Adam); and (iii) the impact of network architecture (VGG-19 versus AlexNet). The results demonstrated that initialization with the content image accelerates convergence and better preserves the original structure. It was also observed that the LBFGS optimizer presents superior final visual quality compared to Adam for this domain, and that the VGG-19 network outperforms AlexNet in extracting complex stylistic features. Finally, a software tool based on the fast transfer method by (JOHNSON et al., 2016) was developed, enabling the real-time application of the technique for academic demonstrations and showcasing the evolution of these algorithms’ efficiency.