After experiencing the 1991 total solar eclipse in Baja California, I realized that although I had enjoyed the eclipse, there is no way my photographs could do it justice, especially as camera problems prevented me from obtaining outer corona exposures. It was then, while still in Baja, that I resolved to make a computer mosaic that could show what the eclipse really looked like in a way that no normal photograph can. Hopefully, this approach would do a better job of capturing the astonishing beauty of the eclipse as it appears to the human eye and depict a full range of coronal detail that does not show up on regular photographs.
After a long and sometimes tedious process, I was able to produce this mosaic of 5 pictures taken from Baja California. Each photograph was properly exposed over a progressive range of solar radii in the corona. Simple photographs properly render either the inner corona or the outer corona, but not both because the extreme variation of brightness with distance from the sun in the corona exceeds the latitude of the film. To get the best possible results, I searched for months for the best photographs I could find (or talk people into letting me use). For the final version, I settled on a sequence of five photographs.
The inner four (shown here) were taken by Dennis di Cicco of Sky and Telescope, using whatever equipment and film. These photographs have exceptional inner and middle coronal detail as well as excellent resolution of the prominences. The outermost picture (also shown) was taken by Gary Emerson of Golden, CO, using a 500mm f/8 telephoto lens with Fujicolor 100 35mm negative film. Gary chose this film because of its relatively wide latitude. This photograph features a wide field, and a deep enough exposure to go all the way to the sky background with a minimum of vignetting. The di Cicco photographs were shot on medium format film, then transferred to 4X5 color transparencies. These became the raw material for use in the eclipse mosaic.
The first step involved digitizing the images. This was graciously done by David Sime of the High Altitude Observatory located at the National Center for Atmospheric Research. He used a Kodak scanner that has a linear CCD array of 4096 pixels. The five photographs were scanned monochromatically with about 2000 by 2000 pixel resolution. It was beneficial to use 12 bit intensities or 4096 different gray levels to resolve faint coronal details. The innermost photograph was also scanned through a red filter to distinguish the prominences, but this turned out to be underexposed and unusable. The scanned images were compensated for a "venetian blind" effect where the brightness of every other row was offset by a constant.
When I first looked at the scanned images, I knew that they had the quality I needed. Very fine detail was visible in the prominences, and the film grains were clearly visible in the outer corona. Compared with the photographs and the computer processing, the scanning process was not to be a significantly limiting factor in the mosaic's quality. In fact, the prominence detail and lunar mountains turned out to have even greater resolution than could be displayed in the final image of 1024 by 1280 pixels.
Each image was precisely registered using prominences as reference points. The moon moved too much relative to the sun between some of the pictures for it to be used. I had to be careful that the prominences themselves did not change too much between exposures. For the outer corona pictures, the prominences became difficult to see and use for registration. Here, the star Delta Geminorum located nearly 1 degree from the sun came to the rescue. Using these guide points, the images were remapped. Because the images were scaled down by as much as a factor of two in the remapping process, the pixel averaging actually increased the signal to noise ratio, gaining an extra bit (factor of two) or so of intensity resolution. A mask was produced for each image based on brightness. Each picture thus contributes to the mosaic in a roughly annular region where it is the most properly exposed. It turned out that combining the coronal brightness range covered by each picture barely produced enough overlap for the mosaic to work. To match the brightness of each picture relative to each other, a two step process was performed. First, for the middle pictures, a contrast stretch was applied to make the radial variation of brightness relatively linear. This required measuring the observed radial variation, fitting a polynomial function to it, then "linearizing" this variation. This helped correct for nonlinearities in the films response. The actual coronal variation in brightness was assumed to be linear enough at this stage, given that the variation was measured along carefully chosen azimuths.
Secondly, each image was given a linear contrast stretch to match the outer brightness (and radial brightness gradient) of the first picture, to the inner brightness and gradient of the second picture out, etc. In other words, the inner corona picture was adjusted towards the bright end and outer corona pictures were adjusted towards the low end to provide a smooth blend from one to the next. This stretch had to be very accurate to prevent the occurence of seams at the picture boundaries; the appropriate stretch parameters were automatically generated at 36 different azimuths around the sun. At last, the well exposed portions of each image could be combined into a mosaic.
The mosaic at this point mimics what would be generated by a CCD camera. This "CCD equivalent" image represents faithfully the full dynamic brightness range of the corona. There is a one-to-one correspondence between true coronal intensity and brightness in the combined image. When the image is displayed on the monitor, very little coronal detail is visible. Uniform contrast compression of the great brightness range of the corona down to the range of the monitor decreases the contrast of local details. One would need some type of special display system with perhaps a 10000-1 dynamic brightness range to display realistically the corona and prominences! The data itself contains several thousand different gray levels, more than either the display monitor or the human eye can resolve. The question is how to display the mosaic using media of limited brightness range such as a computer screen, photographic film, or a magazine reproduction. An important goal was to show the beauty of the eclipse on one image as it truly appears to the human eye using good (about 20x) binoculars, something no ordinary photograph can do. As one might guess, false exaggerated colors were not considered; I wanted to show as much coronal detail as possible without destroying the realism of the image.
What was done instead is to digitally apply a radially dependent linear contrast stretch. This is roughly what a radially graded filter, long used for eclipse photography, does in an analog fashion. However, the digital approach has more flexibility than a radial filter allowing both the brightness and contrast of the image at different solar radii to be optimized after the event. Important information used in tuning this radially dependent stretch was the minimum and maximum brightnesses at various solar radii. This had to be measured on the image at those azimuths having long smooth streamers. After some tinkering, a pleasing image rendering the larger streamers over the full range of solar radii was the result. Some recent refinements taper the coronal brightness very gradually in the outer portions with a much steeper increase in the innermost corona. The relationship between overall radial brightness variation and streamer contrast can be tuned for a particular reproduction medium.
Even though large coronal streamers could now be seen, many of the finer streamers were difficult to pick out. In addition, the film grain was too obvious in the outer corona. Spatial filtering was next applied to the image to address these types of problems. Once again, parameters were allowed to vary smoothly as a function of solar radius. The goal of the filtering was to suppress details of certain sizes (or spatial frequecies) like small (high frequency) film grains and enhance or boost details of medium size or frequency (like smaller coronal streamers) in a manner similar to unsharp masking.
Advantage was taken of the fact that outer coronal detail is preferentially oriented in the radial direction inward and outward from the sun. To do this, separate filters were applied in the radial and azimuthal direction. In the inner corona, medium and high frequencies (small details) were boosted in both the radial and azimuthal direction. Masks were constructed to delineate the moon and prominences to prevent "ringing" from the filter. In the middle corona, medium frequencies were boosted in the azimuthal direction but very high frequencies were suppressed for all orientations. Finally, in the outer corona, middle and high frequencies were suppressed, but much more strongly in the radial direction than in the azimuthal direction. This had the neat effect of letting what would otherwise look like a set of disorganized slightly brighter film grains lying along a radial line from the sun snap into a faint coronal streamer when the filter was turned on. A special localized high pass boost filter was applied to bring out a magnetic loop near the sun's southwest limb. The input photographs, other photographs and photographic mosaics, and my memory of the eclipse all served as guides to check on particular coronal details and other aspects of the mosaic.
Color rendition is one of the more artistic aspects of the mosaic. The trick is to preserve the color balance of the image in the face of locally varying and enhanced image intensities. After much tinkering, lookup tables converting the monochromatic image intensity to separate intensities in red, green, and blue were developed that are also a function of radius (and hence true coronal intensity). The purpose of this was to produce an image exploiting the 24 bit color (16 million different colors) capability that is present. With 8 bit color (256 possible colors), the hue would have had to be a nearly strict function of the intensity of the image. This produces unrealistic results. Although the processed image no longer has a one-to-one relationship between image intensity and "true" intensity, it was desired to reasonably restore one-to-one relationships for hue and color saturation. The upshot is that darker inner corona areas should take on a more truly yellowish gray appearance rather than the blue sky color that is only appropriate in the outer corona. The inner corona was portrayed as mostly white with a slight hint of yellow, I took the color of the sun itself as a guide.
A starting point in producing these color relationships were some color scans of photographs made for an earlier version of the mosaic. These were taken by Timothy J. Brown of Boulder, CO., and Willis Greiner of Conifer, CO. The photographs also helped in many other ways to gain experience with the mosaicing process.
Special processing was done for various elements of the eclipse. The maria on the moon lit by earthshine actually showed up on Gary Emerson's outer corona picture. Though photographed at one or two previous eclipses, the author would like to know if this has ever been visually observed. This region from Gary's image was reregistered to match the location of the moon in Dennis Di Cicco's inner corona image. A separate radially dependent contrast stretch was performed to suppress film halation from the brilliant inner corona, then a filter passing only medium frequencies was applied. The halation with 3' of the moon's limb was too great for detail to be apparent.
The prominences were "broken out" from the rest of the image by comparing the local intensity with neighboring coronal areas. This provided a separate dataset that was manipulated into a prominences only image, the rest of the field being rendered black. The 3.5 magnitude star Delta Geminorum had to be given "protective" windows to prevent the star from being smeared away by the filters that suppress the high frequencies.
Many other smaller details had to be taken care of. These included localized filters to remove ring and arc shaped artifacts in the image, rather like figuring zones of a mirror under the Foucault test. Ring effects caused by the quantized brightness in smoothly varying regions of the outer corona were suppressed by "dithering". This same effect could be noticed in sunset photographs of Mars taken by the Viking Lander. As an example, an area having a brightness value of 55.5 counts in the output image was dithered by randomly giving half the local pixels a value of 55 and the other half 56. A noise filter was developed to recognize and remove defects caused by dust on the transparencies during scanning.
The above processing took place on a VAX 6420 computer. Virtually all of the software to do the image processing was written from scratch. It was considered too difficult to use off-the-shelf software for applying the various image processing operations in a suitably localized and efficient manner. A full image run can be done in several minutes of CPU time. The source code contains about 5000 lines of FORTRAN.
The mosaic was finally displayed on a Stardent computer having a high resolution SONY Trinitron monitor that can show practically the full 1024x1280 pixel resolution of the image. The display also reproduces 24 bit color which means 256 intensity levels in each of the primary colors - red, green, and blue. This allows a total palette of over 16 million colors. The software on this computer, called AVS (Application Visualization System), was very useful in comparing different versions of the mosaic and testing various contrast stretch operations. Finally, a color slide was produced by photographing the computer display screen with Fujichrome Velvia 35mm film that has high resolution and good color saturation. The 256 brightness levels in each color were not sufficient to capture the fiery pink-red prominences simultaneously with the fainter parts of the corona. A double exposure was thus used, superimposing the prominences-only image (using a long exposure to really burn them in) with the regular image that has the prominences subtracted out.