for the candidate selection in (3). With
this improved candidate selection process in combination with
simple greedy inpainting (selection of one replacement patch at
each position), we already obtain a clear improvement over the
earlier method from [7], as is visible in Figure 8 (notice, in particular, that the effect of deleting parts of letters is less severe).
One can also select multiple candidates, with several largest values of P, ij in (3) and subsequently solve the resulting “puzzle”
using a global optimization method like in [17] and [25]. It
would be interesting to explore also alternative solutions, like
the statistics of patch offsets [16] or hierarchical, superresolution-based inpainting [18].