Canon USA's Explanation of
the XL1 Banding/Posterization Problem
as told to Chris Hurd by Joe Bogacz
with Don Palomaki
I first met Joe Bogacz, Tim Smith and Mike Zorich of Canon USA back in November at COMDEX '97, just prior to the release of the XL1 to the prosumer market. Obviously, a lot of water has passed under the bridge in the five months since then and I looked forward to seeing these guys again, even though the circumstances were somewhat different this time. The Watchdog pages had really taken off, appearing in three magazines and linked from numerous websites. Needless to say, when I approached the Canon booth at NAB, these guys were full of smiles and handshakes and they were quite ready to deal with me.
Joe Bogacz, Canon USA's Director of Video Technology, was well prepared and eager to address the issue that's been commonly referred to as the "posterization/banding effect." It was Monday, April 6th, the first day of NAB. "You're about to be one of the first to hear this," he told me as he led my girlfriend Kelly and I into a small presentation room behind the booth's main counter. "The only reason you can't tape this is because I really haven't delivered the speech often enough to feel comfortable, but feel free to take all the notes you want." I sat down with pen and paper and gave him my full attention, stopping him every so often to verify my notes.
"This is either good news or bad news depending on how you look at it," Joe began. "The problem is real and it's there, and it's been around before the XL1 was ever designed. It's part of the DV technology and the proper term for it is called contouring." The implication was simple and carried enormous weight... the good news was that the XL1 isn't suffering from some sort of engineering flaw. The bad news, however, was that the XL1 is showing, under certain circumstances, an anomaly of digital quantization, and for a very good reason.
"When we designed the XL1," Joe explained, "the holy grail that we were striving for was to produce an image as noise-free as possible. The video signal of the XL1 is so clean that unfortunately there are occasions when an image effect known as "contouring," which is a result of the digital quantization process and occurs in all DV cameras, may be visible on the XL1 even though it's masked out on other video cameras by their signal noise."
Joe proceeded to show me a Beta SP videotape on a professional video monitor to demonstrate this process. The first example was a shot of an orange, with a sort of fresh-picked texture to it, against a blue background which gradated in tone across the screen, recorded at -3db gain. The contouring, so called because the bands of color variation resembled the curving lines on a topographical map, was readily apparent. You couldn't miss it, it wasn't subtle by any means. "That's what everyone is talking about," Joe said, as I stared at the screen.
Subsequent portions of the demonstration showed the same scene at increasing gain settings. With each increment of additional gain, the contouring bands became less noticeable. At +12db, the XL1's highest gain setting, they were almost invisible. Almost, but not quite. I could still see a couple of bands of different tones of blue, but the edges of those bands weren't very well defined, plus of course I already knew where to look for them. And although these bands were still discernible on this second-generation Beta SP tape, in all probability they would disappear altogether in the generation loss of going to VHS.
However, I believe that if there should be any change to future XL1's, it would be highly beneficial to increase the user-adjustable gain setting to +18db as on most other cameras. This will most likely completely eliminate the contouring effect. Michael Pappas has noted that in his recent side by side comparison between an XL1 and a Sony VX1000, the XL1's picture at +12db matched the VX1000's picture at -3db. In other words, the noisiest XL1 image was at least as good as the VX1000's cleanest image. That says a lot about the XL1's picture. An option to go to +18db on the XL1 shouldn't significantly alter its image in most cases, and should virtually mask any contouring, which was still just slightly noticeable at the XL1's highest gain setting of +12db.
Joe explained that the contouring effect, a direct result of what's known as a quantizing error in the current DV technology, occurs in all cameras with DSP (digital signal processing). I asked if it was discernible in other cameras and he said it could be seen in several different makes and models and he had an example to demonstrate. We continued viewing the Beta SP tape and I watched the same scene again, this time shot with a Sony BVP-70. Interestingly enough, the contouring bands were indeed visible on the screen, although the picture was noticeably noisier and the edges between the bands weren't as sharply defined.
This quantizing error is discussed in An Introduction to Digital Video by John Watkinson (ISBN 0 240 51380 0) in chapters 2.9 and 2.10 and is caused by the process in which an analog CCD signal is transformed by digital conversion. It's an effect which has been around long before the XL1. The "fix" is to add dithering, which in video terms means adding signal noise, or increasing the user-controlled gain setting (or also, as Joe pointed out, changing the subject background such that the gradiant tones showing the contouring effect are eliminated).
Higher-end cameras with DSP avoid the contouring effect by spending money on it. Contouring is abated in broadcast-professional cameras by using large, expensive CCD blocks; expensive digital signal processors and expensive ADD converters. "For the XL1 to not show contouring under any shooting circumstances," Joe said, "it would have to cost upwards of fifteen thousand dollars, and that's not the market we had in mind."
"For instance," Joe explained, "the XL1 has a nine-bit processor. During our research into this matter, we looked at the possibility of going to a 10-bit DSP but that didn't seem to affect the contouring much. It's entirely possible to switch to a 24-bit DSP like some of the high-end cameras have, and that would certainly eliminate the banding. But the cost of those processors would make the XL1 pretty much unaffordable for the people we intended to have this camera in the first place."
The XL1 is a "prosumer" camcorder. It was designed that way from the start. For the XL1 to cost less than five thousand dollars, a choice was made between enjoying a high Signal-to-Noise ratio (a clean image) which occasionally shows the contouring effect, or suffering from a low Signal-to-Noise ratio (a noisier image) without any contouring. Canon, of course, went with the former, and as stated above, suggests that increasing the gain setting is the best way to deal with the contouring effect if it occurs.
Joe pointed out to me that the amount of background noise is also controlled by the quality of the on-board voltage regulator (for the XL1, it's the best available), the type of CCD block used (three one-third inch 270,000 pixel chips) and the design of the printed circuit board. Once again Joe mentioned the XL1's price range. "This is the cleanest image you'll find in this price range," he said. "But it comes with a trade-off. If you're in a high-dollar shooting situation then you should be shooting with high-dollar equipment. That's always the best insurance." As I've pointed out to many people myself, what Joe was saying is the same thing Canon has insisted all along: the XL1 has many "pro" features, but it is NOT a "pro" camera. However, for the money, and that's the ultimate determining criteria for a lot of folks, it comes very, very close.
Watchdog contributor Don Palomaki adds: "Canon's explanation is not a big surprise, and the FAQ text kind of corroborates (in a indirect way) the thoughts I had about too clean a signal out of the CCD amps, perhaps coupled with some DSP effects that expand the gray scale to produce clear, 6-bit resolution bands.
"Were the noise into the AD converter at, say -10 db below the quantizitation steps), there might be no obvious banding thanks to noise-induced dithering that would soften the band edges. That would correspond to a S/N of very roughly -55 db relative to 100 IRE. But the XL1 S/N is more like 60+ db per the review in Video mag, so the quantizitation band edges become more apparent. By the way the noise may not be off the CCD's, it may be in the amplifiers that are between the CCDs and the AD converter.
"Obviously a quality monitor and an expanse of gray scale is needed to see it clearly. Small monitors and steep gradients provide what I'll call "spatial dithering" by making any bands sufficiently narrow to not be noticed by most observers.
"One problem with a CODEC firmware fix is, of course, distinguishing between gradients (that should be dithered) and real bands (which should not be dithered) in the subject. Wish I had an XL1 to fiddle with the try better quantify the problem. I guess that Canon is banking on the apparent fact that for 90% of potential users and 99% of likely scenes, while present, it is not an issue.
"One test I would try is a high resolution frame grab (use the S-video output) from an XL1 showing with visible banding, and the same scene from a VX1000, maybe use a high quality Snappy or some such beast and save it to a loss-less format, like 24-bit TIFF, BMP, or PCX file (the Watchdog notes: this is essentially the test Michael Pappas performed which convinced him that the contouring was not a camera-borne artifact). Then I would plot the levels down the gradient. I would expect to see a noticeable stair step at each of the XL1 bands. With the VX1000, I would expect to see a hint of a stair step, but with very fuzzy values at the steps. The Canon might go something like:
1 1 1 1 1 1 1 1 1 1 1 2 1 2 1 2 2 2 2 2 2 2 2 2 2 2
... and the VX1000 something like:
1 1 1 1 2 1 1 1 2 1 1 2 1 2 1 2 2 1 2 2 2 1 2 2 2 2
with the XL1 having a sharper band edge."
Thanks, Don. You always seem to explain these things so well.