Files
Abstract
Real-World Facial Super-Resolution (RWFSR) is a complex problem of producing high-resolution face images from low-resolution images captured with real-world image degradation to assist in downstream facial recognition tasks. While many approaches to RWFSR have been developed, there is a lack of hybrid models in the literature that explicitly tackle both unknown degradation estimation and identity preservation. This thesis moves toward a hybrid approach in two experiments: in the first, three state-of-the-art super-resolution algorithms are compared on benchmark training and testing datasets that are representative of simple, complex, and real-world image degradation; in the second, a novel approach is introduced that combines elements of the two best-performing algorithms from the comparison study and is evaluated on benchmark datasets as well as the MILAB-VTF(B) face image dataset. The results from these experiments reveal a trade-off between restoring images with complex degradation and maintaining identity features of face images within super-resolution models.