Colorização automática de imagens em escala de cinza utilizando Deep Learning.

Abstract

Convolutional Neural Networks and other Deep Learning techniques have become more relevant outperforming numerous classic applications in fields like Computer Vision and Signal Processing. Therefore, this project proposes a study covering those techniques in order to ascertain the viability of implementing them in order to colorize grayscale images automatically, something that has been studied by Computer Vision experts since the decade of 1980, and determine the most efficient way of doing it. Three Convolutional Neural Network models, trained to perform automatic colorization, were evaluated according utilizing the root mean square error (RMSE), the peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM) together with an user study, thus allowing the strengths and weaknesses of each model to be determined. It was concluded that it is not only possible to implement automatic colorization utilizing Deep Learning, but results that are good enough to convince other people of the colorization’s authenticity can be achieved.