Style Transfer with Generative Adversarial Networks

This dissertation is focused on trying to use concepts from style transfer and image-to-image translation to address the problem of defogging. Defogging (or dehazing) is the ability to remove fog from an image, restoring it as if the photograph was taken during optimal weather conditions. The task of defogging is of particular interest in many fields, such as surveillance or self driving cars. In this thesis an unpaired approach to defogging is adopted, trying to translate a foggy image to the correspondent clear picture without having pairs of foggy and ground truth haze-free images during training. This approach is particularly significant, due to the difficult of gathering an image collection of exactly the same scenes with and without fog. Many of the models and techniques used in this dissertation already existed in literature, but they are extremely difficult to train, and often it is highly problematic to obtain the desired behavior. Our contribute was a systematic implementative and experimental activity, conducted with the aim of attaining a comprehensive understanding of how these models work, and the role of datasets and training procedures in the final results. We also analyzed metrics and evaluation strategies, in order to seek to assess the quality of the presented model in the most correct and appropriate manner. First, the feasibility of an unpaired approach to defogging was analyzed, using the cycleGAN model. Then, the base model was enhanced with a cycle perceptual loss, inspired by style transfer techniques. Next, the role of the training set was investigated, showing that improving the quality of data is at least as important as the utilization of more powerful models. Finally, our approach is compared with state-of-the art defogging methods, showing that the quality of our results is in line with preexisting approaches, even if our model was trained using unpaired data.