In this dissertation, three example-based single-image super-resolution methods and a benchmark study are presented. The three super-resolution methods individually explore domain-specific, efficient and effective super-resolution solutions.

The first method is developed for face images which contain domain-specific content. Test images are decomposed into facial components, edges, and smooth regions to develop adequate upsampling processes independently. Exemplar regions are exploited to transfer high-resolution details to reconstruct high-quality facial components.

The second method is designed to generate super-resolution results efficiently for generic images. Multiple regression functions are trained to predict high-resolution patch features from low-resolution ones. By splitting the feature space into numerous subspaces and collecting sufficient exemplars for each subspace, the trained regression functions efficiently generate effective features to reconstruct high-resolution images.

The third method integrates regression functions and patch exemplars to fully exploit exemplars to generate high-quality super-resolution images. As regression functions stably estimate high-resolution features and exemplar patches contain rich high-frequency signals, the proposed method uses regression functions to generate a robust intermediate high-resolution image and then finds effective exemplar patches to enrich the high-frequency signals.

The benchmark study systematically compares the performance of state-of-the-art super-resolution methods under numerous parameter settings and test images. It investigates the effect of important parameters qualitatively and quantitatively and figures out the effectiveness of many metrics via human subject studies.

In summary, this dissertation thoroughly and deeply investigates single-image super-resolution problems and propose solutions using exemplar images.