are relatively small compared to the crack width, it is very difficult
for the inpainting algorithm to infer the correct structure locally.
Patch-based inpainting methods typically handle structure propa-
gation by defining the right filling order [9], [15]. Once the filling
order is determined, most of the methods choose plausible candi-
dates for replacement patches based solely on the agreement with
the undamaged part of a single target patch [see Figure 7(a)], and
concentrate on defining effective distance metrics between the
known portion of the target and the candidates [20]. Matching only
against a small part within the target patch increases the risk of
propagating wrong textures and wrong colors into the missing
region. Global methods, such as [17] and [25], allow multiple can-
didates and optimize their mutual agreement in the overlap
regions [Figure 7(b)], but even this cannot ensure agreement with
surrounding undamaged structures: the optimization that takes
care that neighboring replacement patches mutually agree cannot
“undo” the damage done by selecting wrong candidates in the first
place. Ideally, undamaged areas around the target patch should be
taken into account in the candidate selection as well, ensuring that
plausible candidates agree with true structures [see Figure 7(c)].