Pre- and Post-processing on Generative Adversarial Networks for Old Photos Restoration: A Case Study

Abstract

Old historical images are an invaluable source of knowledge that allows people to learn about past events and, in general, the form of the world in the past. In the case of townscapes, the photos may depict specific details as building appearance prior to their reconstruction, enlargement or demolition, or even former appearance of cities (buildings, inhabitants, transportation, among others). In this sense, more and better details of the image lead to an exact representation of a city in a given time. Generative Adversarial Networks (GANs) are a category of deep artificial neural networks (DANNs) that show great success in generating realistic characteristics into image, video and voice data. This work explores how the pre- and post-processing techniques influence the overall effectiveness of GANs-based techniques for restoring and coloring old photos. Pre- and post-processing based on traditional image processing methods preserve and enhance the information contained in old photographs; however, their effectiveness is limited by the amount of information retained in the original photograph. On the other hand, GANs-based techniques offer the ability to increase the amount of information and thus boost the effectiveness of traditional methods. Experiments are performed referring to the old photos of Quito’s city. The preliminary results show that pre- and post-processing algorithms are essential even in artificial intelligence approaches, eliminating undesirables artifacts and increasing visual quality.