Super Resolution and Deblurring of Hyperpanoramic Images Using Neural Networks

In this thesis we studied the application of Deep Neural Networks to the problem of Super Resolution of images, which consists in enhancing their spatial resolution and perceptual quality; we focused on images captured by a hyperpanoramic lens which presents spatially variant blur caused by its point spread function. We investigated two different architectures, CNNs and Transformers, together with specific network designs particular for the task of Super Resolution such as adversarial networks, and variants of the Adam optimization algorithm for training of deep networks. We modified an existent pipeline for generating supervised training data to let our networks learn the spatially variant degradations of the panoramic images, and evaluated their performance with a no-reference image quality metric. Both architectures perform well in Super Resolution, with Transformers generally outperforming CNNs; by attribution analysis, we understood that Transformers have in fact a larger receptive field. We also found that adversarial training, while improving image quality, is more susceptible to the generation of artifacts.