Learning super-resolution from misaligned examples

Implementations of Example-Based Super-Resolution (EBSR) have been developed extensively. Any such EB-SR method is typically evaluated against a constructed test set as to define its performance and applicability. Nevertheless, it is rare for a formed test set to precisely resemble data met in a real-world problem. Usually, low-quality training and test subsets are obtained directly from their corresponding high-quality ground truth data. This allows for a complete and reliable quantitative examination of performance at a later stage. In a real-world problem however, test data are obtained from another source, as for example, printed images. Naturally, low-quality scanned halftones and high-quality continuous tone images would possibly be spatially incoherent training pairs. Such circumstances give rise to one major consideration, misalignment in training subsets. The present work demonstrates the significance of effect of misalignment among training subsets in applying EB-SR and supports the necessity of image registration in preprocessing to overcome this problem.