Contents
Content on this site and text of this page Copyright 2004 Tom Fiddaman. Images on this page are in the public domain. Link freely.
|
|
As an early Sony 828 adopter, I've been following the CA/purple fringing debates with considerable interest, though to be honest I've only seen really bad fringing on two shots out of 600 or so (and those weren't keepers anyway). It has been a minor annoyance in a few others, though not noticeable in prints (I can already hear the "blind MR" flames starting). Also, this is Seattle, and I'm a little worried that it'll be a problem when the sun finally comes out.
Disclaimer: I have little prior expertise with optics (apart from photography) but have no shortage of opinions on things I know little about.
I've seen a bunch of theories about its cause:
(Note that there's some variation in use of terms in 1 and 2, so if I have it wrong, read as I mean not as I say).
If it's 1, each channel should have a sharp image, but offset a little from the others. The PanoTools trick
of rescaling the blue channel (i.e. effectively changing its magnification) should work (it doesn't work well for
me).
If it's 2, some channels should be sharp and others fuzzy, causing a halo in the fuzzy channel. The restoration
problem is hard - you'd somehow have to restore the blue channel to focus by regularized deconvolution, or perhaps
cheat and guess at the restored fuzzy channel by assuming it covaries with the "good" channel. Very painful
either way.
If it's 3, there might be some hope for restoration (especially if Sony would open up the encrypted RAW format).
But I seriously doubt that it is, as the extent of the aberrations would then be limited to 2-pixel blocks.
If it's 4 or 5, presumably the extra UV or IR would be unfocused, polluting one of the other channels and causing haloes as in 2. Easy to correct with a filter. I've had a UV0 on from day one and still get PF so this is not it.
6 should also cause haloes, but given the Bayer pattern they should spill over into other channels as well (i.e. an overexposed blue sensor leaks charge into adjacent R and G sensors).
There's a useful page on CA at Nikon MicroscopyU.
To get to the bottom of things I created a CA/PF torture test and shot it at wide, medium, and tele (yes, with a tripod). I neglected to white balance but it shouldn't matter. I used ISO 64 and default settings except low sharpening. I found the following:
- At wide angle, a fair amount of PF, with fringes near 10 pixels wide at the corner, but nonexistent near the center
- At medium and tele, next to nothing visible, except around highlights on the glossy paper when I inadvertently left the flash on
I think the lack of issues near the center of the field partially rules out theory 2, unless different colors have different field curvatures.
To see what was really going on around the fringes, I borrowed a plugin I'd written (using FilterMeister) to plot RGB values across a slice of an image. It takes a selection, dims it (for visibility of the plot), adds a horizontal or vertical centerline, and then plots the RGB values along that centerline. This makes it easy to get a visual sense of noise, edge fidelity, etc. Right away I saw weird things. The following is a crop from the bottom right corner:
If you look at the side of the plotted area, where the edge of the paper meets the floor, you can see the purple fringe. At that edge the red and green channels are close to ideal step edges (though there's some overall softness - this is 28mm equivalent, F2 at close focus in macro mode so it's worst-case for the lens). Interestingly, they overshoot somewhat, as you'd expect from sharpening, but on a 5-10 pixel scale - too big to come from the camera algorithm. The blue channel is equally sharp approaching the edge from the left, but then tapers off rather more slowly. It's this excess blue that appears to be responsible for the purple cast. Elsewhere, you can see the same asymmetry - perfect edges from the left, and some blur or blooming or whatever to the right.
Right away I think this eliminates ordinary CA (theory 1 or 2) from the running as neither form predicts this asymmetry.
One other surprising thing to note here is that there are a few segments where all three channels are perfectly noiseless for a few pixels. I think this may be due to anomalously clean 8-pixel JPEG blocks. It not, it suggests to me that the chip's noise issues are not photon noise but other things. A little care is in order as the paper used was not completely texture-free.
For the record, here's the center of the field from the same exposure as above:
My first test pattern was a little busy, so I created a second. The following is a crop from the top right of the frame (same shooting conditions as above):
Notice that there's progressively less fringing as you move toward the center axis (off the image, bottom left). Plotting the edge values yields about the same results as above - ideal edges from the left, but blurring to the right. Looking at just the blue channel (gamma adjusted for improved visibility):
After some experimentation with PanoTools-type fixes (rescaling the blue channel) I realized that the blue channel is neither out of focus nor out of scale. Instead, there's a primary blue image that is sharp and aligned with the R and G channels, plus a secondary ghost image displaced about 10 pixels radially outward at the corner, with intensity about 1/3 that of the primary image.
The good news is that this is easy to fix. Generally correcting lens aberrations is a hard inverse problem. But in this case it's easy, as the PF doesn't pollute the image that badly. Thus you can use the degraded image as a starting point to simulate some more PF, then subtract the simulated PF and get back to something close to a perfect original. In this case, subtracting blue at 90/255 intensity, displaced radially inward 9 pixels at the edge and 0 at the center, produces the following:
Note that the correction of the bluish fringe is near perfect. Theres a little bit of residual reddish-green cast in the black areas where the purple was removed, and some residual yellow at the inner white borders. I suspect the latter is correctable by playing with the R and G channels as well, or is conventional CA, or perhaps a sharpening or interpolation artifact. Either way, it's only visible in high-contrast 100% crops so who cares.
Following these tests I was pretty confident that PF could simply be erased from my real world images. Not so fast. First, consider the following tree with purple raindrops:
 Original, 100% crop from top center of image. These are small fringes but surprisingly visible even when the image is viewed at much lower resolutions. I'm not sure I should care, but I'm annoyed in principle. |
|
Notice that the fringe is on the inside this time (as in the test targets above, the center is unpolluted). This image was shot at 200mm (equiv) F2.8 - obviously something about the lens geometry has changed. It's also harder to correct, as the raindrop highlights are blown out so the maximum blue channel value of 255 is a poor estimate of the true intensity that created the fringe. It's possible to improve the fix somewhat by guessing a uniform value >255 for highlights, but that introduces color artifacts. Admittedly they are minor, and by overcorrecting (more than shown above) it's possible to eliminate the appearance of purple at lower magnifications without disrupting the overall integrity.
Both fixes slightly change the overall color balance of the image, in a way that I find pleasing (probably hard to tell from this tiny crop). This agrees with the contention that, while PF is only visible in extreme conditions, it is in fact present everywhere. Hard to say for sure without much more careful testing.
Unfortunately, things get even worse. The following are crops from an image of (overexposed) vine maple leaves against a dark forest background.
|
100% crop, top right corner |
100% crop, center. |
These differ from the prior images in several ways:
Looking at the individual channels is helpful:
Blue channel only; blown highlights (B=255) shown in yellow, a few lowlights (B=0) in cyan. It's clear that the blue fringe on the thick stem is present only where the blue sky was overexposed. The blue component of the purple fringe on the thin vertical stem and leaf surrounds those bright objects. In spite of the fringe, the boundary between the thick stem and the overexposed highlights is sharp, which suggests that the blue channel is not misfocused here. Interestingly, the fringe is less intense at the edge of the stem; if I had to guess I'd say that's a sharpening overshoot. |
 Red channel only; highlights in yellow as at left. Blocks are JPEG artifacts from intermediate saves of the file (oops). Since the sky at bottom is not overexposed, there's no red component against the thick stem, which is why those fringes are pure blue. On the vertical stem and leaf, there is some overexposure, which is apparently responsible for the red component of the PF. The green channel is relatively clean. |
So far I think there are two things going on.
First, there's an obvious source of fringing from a sharp but radially offset secondary image in the blue channel. It's geometry varies - sometimes the fringe is radially outward, sometimes inward, and it seems to be worse at wide angle and wide open. If this is some well-known lens phenomenon that I'm ignorant of, I hope someone will enlighten me, but it doesn't fit my (admittedly sketchy) understanding of conventional chromatic aberration. If I had to guess, I'd pin it on some insufficiently multicoated object in the optical path (the hot mirror?) that is creating the ghost image via an internal reflection that slightly changes the optical path. It might be possible to test this by shooting a high-contrast target in nightshot mode, but I haven't tried. This source is highly correctable, but blown out blue highlights complicate things by making it impossible to know the intensity of the source of the fringe.
Second, there's another behavior that appears to be a problem around dark objects adjacent to blown highlights. It has a consistent width, no particular orientation and isn't sharp, which is why I think it's different from the first source. It's mostly evident in the blue channel but also occurs or spills over into the red and green channels. I'm unsure about the origin, though I don't think it's ordinary CA. I don't think there's an easy correction in principle, but I'm guessing that it can be defeated with some combination of informed desaturation (exploiting RGB correlations and working only near highlights, not just globally desaturating the magentas) and guessing at blown highly values from their spatial extent.
In both cases, the fringe appears to be a secondary effect of some sort. The primary image in the blue channel is geometrically aligned with the red and green channels and in equally sharp focus.
I don't plan to run many more tests (though I'll certainly be watching for them), but I will continue to refine a two-pronged approach to a universal fix for both types of PF. At the moment I have no commercial ambitions, so at some point when I get some time I'll probably release free copies of the plugins I'm working on. No promises though. I'd really rather see Sony take notice and fix this!
I'm eager to learn more about this, but also rather overwhelmed at the moment, so rather than emailing me, please contribute to the Sony forum discussion at DPreview.
