Generative Adversarial Network-Based Tattoo Enhancement.

Throughout history, visual art has been transforming from primitive cave painting to photorealistic imagery. However, out of the existing variety of art forms, some of them have been undeservedly underrated, such as tattooing. Over centuries tattoos have been considered as stigma due to their affiliation with marginalized communities, only to become a widespread self-expression marking and a fashionable accessory later. In the recent years, with rapid technological developments, the innovative minds have managed to breathe some new life into tattoos, making them the base of biomedical, fitness and remote control devices, as well as temporary tattoos for makeover lovers. Their designs, however, can be enhanced with the use of modern machine learning methods, mainly Generative Adversarial Networks (GAN), which can speed up the production process and generate unique customized arts. Hence, the aim of the Master’s Thesis was to automate the process of black and white tattoo sketches’ colorization with the help of machine learning methods. A sample of 800 images of neotraditional animal tattoos was collected and preprocessed in order to train three GAN models (pix2pix, CycleGAN, DeblurGAN), capable of colorizing a sketch. The results of the final model implementation on the test set were shown to a group of respondents whose task was to evaluate the quality of obtained images and give their feedback. In the end, there was not a clear favourite among the participants of the survey, as the trained models generated three defined beautiful aesthetics, each appealing to different parts of the group. Thus, it can be stated that the model outputs can be successfully used as a variety of options for customized tattoo design production.