Super-resolution with significant illumination change

In this paper, we study the impact of illumination change on super-resolution. Based on analysis and experimental results, we show that large illumination change, not a global/parametric one, poses a significant challenge to existing super-resolution algorithms. Existing robust techniques of handling outliers do not offer a good solution either. We propose a two-step framework to super-resolve the object shape at high-resolution first and then synthesize high-resolution images. For applications where appropriate illumination models are not available, we propose a pre-processing framework. Extensive experimental results are presented based on synthetic and real examples.