Peasant girls. Sergei Mikhailovich Prokudin-Gorskii, ca. 1909.
All images are digitized | All jpegs/tiffs display outside Library of Congress | View All
Reconstructing Prokudin-Gorskii's Color Photography in Software
by Blaise AgŸera y Arcas
Originally, the Prokudin-Gorskii plates would have been viewed using a projector with three lenses, each projecting colored light, produced using the same filters as those used to photograph the corresponding exposures. Blue, green and red light would have mixed on the projection screen to produce full-color images. Today, we have the means to combine the three exposures by computer, producing full-color digital reconstructions of Prokudin-Gorskii's photographs. This can be done by hand, using software like Adobe Photoshop to locate corresponding points in the three exposures, overlay the exposures so that these points match, and render each exposure in a different color (see the section of the Color Photography Method page that discusses Digital Color Renderings). Achieving the correct alignment of the three color components by hand, however, is very difficult and time-consuming. The exposures may be slightly offset or tilted relative to each other, and occasionally they even appear to show subtly different perspectives on the scene. In some cases the emulsion appears to have warped slightly, further complicating alignment.
An alternative to hand-alignment is to develop computational methods to align the color components automatically. The software used to produce color composites of the Librry of Congress Prokudin-Gorskii photographs takes advantage of newly developed algorithms originally designed to align digital photographs of pages from different copies of early books for bibliographic analysis. In addition to the obvious advantages of automation and speed, "hands-free" alignment using software is a mechanical procedure involving no subjective artistic decisions about individual photographs. No cropping, color-balancing, retouching, repair or other manual intervention has been attempted; the color renditions reflect the real-life imperfections of the original plates.
Factoring in Exposure Times
For the majority of the Prokudin-Gorskii images, perfect alignment is impossible to achieve using conventional methods. This is because the three color exposures were not all taken at once, but in the sequence blue, green, red.
Exposure times were quite long, since the photographic emulsion was "slow" and artificial lighting was scarce; for indoor shots, the camera shutter may have been open ten minutes for each exposure, or a total of ? hour for a single color photo. This explains why we find virtually no human or animal subjects among the interior photos. Live subjects were photographed outside in bright light, but even so, they could not in general have held perfectly still for all three exposures, which would have taken at least a few seconds, and perhaps as much as a minute or more. Figure 1 demonstrates the time gap between exposures of the plates.
 |
| Figure 1. Given the radius
of the moon, its distance from the Earth, and
its period of revolution, it can be calculated
that the moon moves across the Earth's sky by
its own diameter in about 2.1 minutes. The moon
is visible in several of the Prokudin-Gorskii
plates. The example above is especially clear.
The animation alternates between "warped" color
reconstructions based on the green exposure and
the red exposure. Between the two exposures,
the moon moves by more than 1/6 of its
diameter, so it would appear that each exposure
took upward of 20 seconds; the whole photo
would have taken over a minute. From the
brightness of the sky and the fairly short
shadows, we can assume that this photo was
taken during broad daylight. On an overcast
day, the exposure would need to have been
longer; indoors, much longer. Small town of Vokhnovo... LC-DIG-prok-21011 |
The subjects of individual portraits clearly attempted to "freeze" (with varying success), but group portraits and scenes involving animals or bystanders often show color ghosting where the scene changed between exposures. In the most extreme cases a subject may be present in one exposure, and entirely absent from another (see figure 2).
 Figure 2. Color anomalies abound in a relatively uncontrolled crowd scene Baiga. Samarkand. LC-DIG-prok-21767 |
Trees waving in the breeze, smoke rising from a chimney, clouds drifting across the sky, flowing water, and many other common landscape elements likewise cause ghosting (see figures 3-4). Therefore, aside from tightly-controlled interior still life tableaux, it is not in general possible to color-align every part of the image at once.
 Figure 3. Color ghosting in the clouds on a windy day. General view of the Church of Saint John Chrysostom... LC-DIG-prok-21229 |
 Figure 4. Moving whitewater yields colorful effects. Study near the Kivach waterfall. LC-DIG-prok-20253 |
The Algorithms
The color alignment algorithm works on pairs of images at a time- the blue and red color channels, for example. It assumes that one of these images, the "target", is correctly oriented; the other image, the "source", is then deformed to match it. The source and target are overlaid, and a grid of small circular windows are defined on these two images. The algorithm then calculates the shift that best maps each region on one image into the other image.
Rigid Alignment
The grid of displacements, or "motion field", can be used in two ways. One is to calculate the optimal single "rigid motion" which approximates the motion field. Rigid motions may also include overall shifts in the plane, stretching and shearing. By using optimal rigid alignment to bring two color exposures into registration with the third (the green exposure is used as the target), a color rendition is produced in which the exposures are aligned to each other as well as would have been possible using Prokudin-Gorskii's projector. In practice the computer-generated rigid alignments are probably much better, in most cases, than the alignments Prokudin-Gorskii could have achieved in one of his slideshow lectures.
Unfortunately, as with virtually all machine vision algorithms, this algorithm may produce errors. When the window used to calculate a shift in the motion field contains no useful information--such as a large stretch of featureless sky or water--then the computed motion in that area is spurious. If there are too many areas like this, the rigid alignment may fail. Alignment also fails when the original glass plate is heavily damaged, with large patches of emulsion missing. Fortunately, the majority of the photos of special historical or aesthetic interest align correctly.
Warpfield Alignment
One can make more use of the motion field by producing three "warped" color images. For each of the three warped images, one of the color exposures is assumed to be in the correct orientation (the target), and the other two exposures are deformed, like sheets of taffy, using their calculated motion fields relative to the target.
-
Figures 5-6 show an example of one of the motion fields used to do the deforming (this is the green motion field, with red as the target). The motion field clearly shows the prisoners moving between exposures. After the appropriate deformations, the three warped color images are something like a three-frame color movie. As the motion field allows us to take small local movements into account, it usually allows a sharper and more precise reconstruction than the color image generated using rigid alignment.
Figure 5. Two prisoners in shackles.
LC-DIG-prok-20009
Figure 6. Warpfield illustration of the
subject movement between exposures. - Figure 7 shows, in sequence, the color ghosting of one prisoner's turban in the rigid alignment, and the disappearance of most of this color ghosting when warping to the blue, green and red exposures is used.
 Figure 7. Detail showing the range of movement of the prisoner's head. Two prisoners in shackles--detail. LC-DIG-prok-20009 |
Historically, while the rigid alignment is a computer reconstruction of what Prokudin-Gorskii would have been able to show his audiences at the beginning of the 20th century, the three warped images are computer reconstructions of the actual scene in front of the camera at the moments when each of the exposures was taken. Although the warped reconstructions allow for much more precise color registration than the rigidly-aligned reconstruction, they are also somewhat less robust, and can produce spurious results. The warped reconstruction composites presented in the online catalog use the green exposure as the target image.
