By Lina Sorg
Art historians and conservators frequently use X-ray radiographs to glean information about not only a painting’s surface, but the cracks and layers underneath. Such techniques also facilitate preservation and restoration efforts in dated artwork. Until the 1950s, paintings done on wood panels—typically from the 1300s to the 1800s—were preserved with a technique called cradling. Cradling involves the attachment of a permanent grid of rectangular or T-shaped hardwood slats to the back of a painting. Cradles was meant to provide reinforcement, uphold the painting’s surface, maintain shape, prevent structural or insect damage, or straighten a warped panel.
Unfortunately, due to their thickness, cradle slats appear on X-rays as visible bright grids that block out the original painting’s features and hinder its study. Art conservators have attempted a variety of methods to remove the presence cradles from X-rays; the current best practice is a time-consuming manual removal from high-resolution digital scans via Adobe Photoshop. This process usually takes between 40 minutes and several hours for one section of a painting. And even with painstaking attention, both the cradle’s textured wood grain and the edges of the cradle slats remain visible in X-rays. Conservators have also used VIPS, an open source image processing software that blends the cradle edges with the painting. But VIPS does not automatically process the space where cradled and uncradled areas meet, and does not treat the cradle’s wooden texture.
In a paper recently published in the SIAM Journal on Imaging Sciences, Rujie Yin, Bruno Cornelis, Gabor Fodor, Noelle Ocon, David Dunson, and Ingrid Daubechies present an algorithm that eliminates the visual discrepancies of cradling in X-rays. The algorithm—and its accompanying multiplicative model—can distinguish, isolate, and remove superfluous textures caused by cradling while leaving the painting’s original textures intact.
Yin et al. developed the multiplicative model, based on the physics of X-ray imaging, after encountering limitations with their original additive model. The updated algorithm works in three phases. First, it automatically detects the cradle’s location and corrects the X-ray’s grayscale inconsistency, resulting from the cradle’s thickness. It also corrects the smooth intensity difference, consisting of discrepancies occurring along the cradle’s slats. The authors then use morphological component analysis to separate the textured grain of both the cradling and the panel from the rest of the X-ray before applying a Bayesian factor model that isolates and subsequently removes the cradle texture.
Yin et al. apply their algorithm to multiple historical paintings, such as Van Orley’s The Pentecost (Figure 1) and an altarpiece painted by Francescuccio Ghissi (Figure 2), and evaluate the results against other methods of cradle removal. When compared to manual removal in Photoshop, their model significantly reduced the cradle wood grain and better handled the intensity difference by producing smoother transitions between the painting and the edges of the cradle slats. VIPS software was unable to isolate and remove the cradle’s wood grain from the X-ray, and did not correct the grayscale difference as cleanly as Yin et al.’s methods.
Ultimately, Yin et al.’s algorithm offers significant visual improvement in cradle removal from high-resolution X-rays. It yields sharper results without compromising the integrity of the original painting. The algorithm also has the benefit of being almost completely automatic, thus eliminating the tediousness associated with manual removal.
The authors have made their methods available to the art community via a tool accessible in both MATLAB and C++. By improving readability of X-ray images of paintings, their algorithm has the potential to greatly improve restoration and conservation efforts for five centuries of artwork.
Read the full version of this article now, and view other SIAM Nuggets.
Source article: Removing Cradle Artifacts in X-Ray Images of Paintings. SIAM Journal on Imaging Sciences, 9(3), 1247-1272. (Online publish date: August 30, 2016).
About the authors: Rujie Yin is a Ph.D. student and teaching assistant in the Department of Mathematics at Duke University. Bruno Cornelis and Gabor Fodor are Ph.D. students in the Department of Electronics and Informatics (ETRO) at Vrije Universiteit Brussel. Noelle Ocon is the associate conservator of paintings at the North Carolina Museum of Art. David Dunson is the Arts and Sciences Distinguished Professor of Statistics at Duke University. Ingrid Daubechies is James B. Duke Professor of Mathematics and Electrical and Computer Engineering at Duke University.