Supervisors: Dr. Radu Timofte
Most of the single image super-resolution proposed solutions assume: 1) the availability of low and corresponding high resolution image pairs for learning, 2) the test images are sampled from the same source (low resolution) image domain as the training images. However, in practice while the domain can be guaranteed if the generation process or the source of the images is known, the training pairs are a luxury and often impossible to obtain. In this work we propose an image super-resolution solution capable to estimate a higher resolution image in the source domain of the image or in another target domain without the need of paired images. We successfully validate our approach on standard and synthesized datasets and compare favorable with the state-of-the-art.