In cases where high-resolution scans are available, such that the
width of some cracks spans multiple pixels, patch-based inpaint-
ing methods [9], [17], [24] typically yield better results [7], [27].
The book in the panel Virgin Annunciate (Figure 2) is a very
challenging case for virtual inpainting because the width of cracks
varies greatly and some cracks are difficult to distinguish from
parts of the letters. Moreover, as the cracks are typically sur-
rounded by bright clouds of background matching color, due to the
lifting and abrasion of the surrounding paint and thereby imposed
light reflections, the immediate areas around the cracks are also
unreliable. The crack inpainting method from [7] specifically tai-
lored to this application already generally improved the legibility of
the text and was shown to outperform some “general-purpose”
patch-based inpainting methods like [9] and [17]. However, some problems are still present, e.g., parts of the letters through which
wide cracks are passing are occasionally deleted after virtual
inpainting. To alleviate this problem, a better approach to handling
continuation of image structures is needed. We discuss a possible
solution with encouraging initial results next.