I am going to make a new set of Neat Image noise profiles for my Canon 10D.
I invite comments about the new methodology.
The new set will use two test targets. The Macbeth ColorChecker card and the Kodak Q-14 Gray Scale.
Why two targets? The Macbeth ColorChecker card has grayscale patches, but those patches are not white or black. They are very light and very dark gray. The dynamic range of the Kodak Q-14 Gray Scale is broader. It goes from 8,8,8 to 247, 247, 247 (no paper is a perfect 255,255,255 and no black ink is a perfect 0,0,0). The idea is to get as much dynamic range for the shadows and highlights as possible, since they show the most noise.
I'm uncomfortable with the profile target offered here. First, it relies on the user to print it. Metamerism, improper profiles, the use of multiple colors to generate grayscales, etc. introduce considerable inaccuracies. Second, there are patches for 0,0,0 and 255,255,255 and those are simply not possible. The Kodak Q-14 is much more realistic, at 3% and 97% gray. Third, the profile only considers grayscale tones. There are no typical colors or saturated colors. Noise is far from uniformly distributed across the color channels in a CCD or CMOS imager. In LAB, the b channel is often noisier. In RGB, the blue channel is often the noisiest.
Each profile image will combine a cropped image of each test target. The images will be taken in the same light and slightly out of focus, an out of focus image is preferable to applying a Gaussian Blur later.
One profile image will be made for each ISO (100 to 3200) for RAW and JPEG (sRGB for the JPEG) for each white balance setting. The white balance settings will include Sunny Daylight, AM/PM Daylight (for warmer light of morning and afternoon), Cloudy, Shade, Photoflood Tungsten, Photoflood Daylight, Fluorescent, and Flash.
Sound reasonable?
Cheers,
Mitch
Methodology for Canon 10D Profiles
Mitch,
It looks like you assume the calibration target should provide some perfect color and brightness values for all users. It should not. Neat Image does not need patches of some exact set of colors or something to build a profile. It is designed to work with what it is available in the image, whether it is your working image or a special shot of the calibration target. That's why the calibration target provided at Neat Image webpage is sufficient (we would prepare another one if it was not). It covers all range of brightness values and enables building and fine-tuning a rough noise profile. BTW, automatic fine-tuning works even if some of the values are missing.
Using standard targets like Macbeth ColorChecker card and the Kodak Q-14 Gray Scale is possible too and may be preferred if you already have one of them and don't want to print Neat Image calibration target. Otherwise, there is no difference.
Regarding profiling, yes, each ISO rate should be profiled separately because ISO rate significantly affects noise levels. White balance has secondary importance in this sense. Usually it is not necessary to build profiles for every WB, but if you want to achieve a bit higher accuracy - do it.
Hope this helps.
Vlad
It looks like you assume the calibration target should provide some perfect color and brightness values for all users. It should not. Neat Image does not need patches of some exact set of colors or something to build a profile. It is designed to work with what it is available in the image, whether it is your working image or a special shot of the calibration target. That's why the calibration target provided at Neat Image webpage is sufficient (we would prepare another one if it was not). It covers all range of brightness values and enables building and fine-tuning a rough noise profile. BTW, automatic fine-tuning works even if some of the values are missing.
Using standard targets like Macbeth ColorChecker card and the Kodak Q-14 Gray Scale is possible too and may be preferred if you already have one of them and don't want to print Neat Image calibration target. Otherwise, there is no difference.
Regarding profiling, yes, each ISO rate should be profiled separately because ISO rate significantly affects noise levels. White balance has secondary importance in this sense. Usually it is not necessary to build profiles for every WB, but if you want to achieve a bit higher accuracy - do it.
Hope this helps.
Vlad
A Bit Easier
Actually, for the RAW images, there is no need to use separate WB settings, since WB is applied by the RAW converter. All that's added to the RAW file is a metadata tag for the WB setting.
JPEGs are another matter. Their WB is applied by the camera.
I do not understand how a set of neutral patches can adequately profile the noise patterns of the Canon 10D. If I convert an image to LAB, the b channel will show more noise than the a channel. In RGB, the Blue channel will tend to be noisy. Images simply do not have equal noise in the three RGB channels.
I want to remove chromatic noise as well as luminosity noise. I don't see how profiling neutral patches will achieve that.
I'm open to being educated, Vlad.
I have both color targets. I use the Macbeth ColorChecker to make ICC profiles for my camera.
Cheers,
Mitch
JPEGs are another matter. Their WB is applied by the camera.
I do not understand how a set of neutral patches can adequately profile the noise patterns of the Canon 10D. If I convert an image to LAB, the b channel will show more noise than the a channel. In RGB, the Blue channel will tend to be noisy. Images simply do not have equal noise in the three RGB channels.
I want to remove chromatic noise as well as luminosity noise. I don't see how profiling neutral patches will achieve that.
I'm open to being educated, Vlad.
I have both color targets. I use the Macbeth ColorChecker to make ICC profiles for my camera.
Cheers,
Mitch
Perfect Colors
No, Vlad. I don't assume that NeatImage needs perfect colors. I'm well aware that noise profiles are different from color management profiles.
The advantage of the Macbeth ColorChecker is that it has patches for typical colors like light and dark skin tones, sky, and foilage. It also has highly saturated colors. So saturated, some are out of gamut for sRGB or ColorMatch.
The reason for using the Macbeth ColorChecker is the combination of typical colors and saturated colors, which provide noise patterns when the colors are not neutral -- i.e., composed with differing ratios of the red, green, and blue.
Now, if you think my theory is flawed, I'm open to persuasion.
Cheers,
Mitch
The advantage of the Macbeth ColorChecker is that it has patches for typical colors like light and dark skin tones, sky, and foilage. It also has highly saturated colors. So saturated, some are out of gamut for sRGB or ColorMatch.
The reason for using the Macbeth ColorChecker is the combination of typical colors and saturated colors, which provide noise patterns when the colors are not neutral -- i.e., composed with differing ratios of the red, green, and blue.
Now, if you think my theory is flawed, I'm open to persuasion.
Cheers,
Mitch
Neutral Target
I'm not persuaded that self-printing will get you even close to a full range of highlights to shadows.
I'll leave the issues of metamerism, paper white balance, and non-neutral grayscale tones aside and point out one issue I think is especially pertinent. Most people do not come close to getting 0,0,0 to 255,255,255 in their actual printed output.
I print on fine art paper with the Epson 2200. I can get a DMax of about 1.8 on something like Arches Infinity paper and black only ink with a matte black ink cartridge, which can hold a lot of ink. Most people cannot come close to a DMax of 1.8. More like something like 1.5. So, your grayscale patches, if you check them with a densitometer, will be something like 20,20,20 and 245,245,245.
The Koadk Q-13 or Q-14 grayscale targets should give considerably more range than self-printing a set of neutral patches on a typical desktop printer.
Cheers,
Mitch
I'll leave the issues of metamerism, paper white balance, and non-neutral grayscale tones aside and point out one issue I think is especially pertinent. Most people do not come close to getting 0,0,0 to 255,255,255 in their actual printed output.
I print on fine art paper with the Epson 2200. I can get a DMax of about 1.8 on something like Arches Infinity paper and black only ink with a matte black ink cartridge, which can hold a lot of ink. Most people cannot come close to a DMax of 1.8. More like something like 1.5. So, your grayscale patches, if you check them with a densitometer, will be something like 20,20,20 and 245,245,245.
The Koadk Q-13 or Q-14 grayscale targets should give considerably more range than self-printing a set of neutral patches on a typical desktop printer.
Cheers,
Mitch
Mitch,
Do you think grayscale patches contain only luminosity noise?
- Neat Image receives enough information about both luminance and chrominance components of the noise from the grayscale patches.
Vlad
Do you think grayscale patches contain only luminosity noise?
- Neat Image receives enough information about both luminance and chrominance components of the noise from the grayscale patches.
Vlad
Last edited by NITeam on Mon Dec 01, 2003 11:20 am, edited 1 time in total.
Mitch,
If a user forgets to measure the white patch during fine-tuning, what do you think Neat Image should do with noise in near-white range?
Neat Image interpolates missing values so that overall noise level vs brightness range graph would be continuous. However, in most cases this is not needed when the calibration target is used and automatic fine-tuning is aplied. You can test this yourself. - Interpolated values will receive yellow shadings, while measured ones - green shadings, as explained in the user guide. In most cases, the calibration target provides enough information to cover all range of brightness.
Vlad
If a user forgets to measure the white patch during fine-tuning, what do you think Neat Image should do with noise in near-white range?
Neat Image interpolates missing values so that overall noise level vs brightness range graph would be continuous. However, in most cases this is not needed when the calibration target is used and automatic fine-tuning is aplied. You can test this yourself. - Interpolated values will receive yellow shadings, while measured ones - green shadings, as explained in the user guide. In most cases, the calibration target provides enough information to cover all range of brightness.
Vlad
I sort of resent this and your previous message, which was also tinged with insulting references.NITeam wrote:Mitch,
Do you think grayscale patches contain only luminosity noise?
- Neat Image receives enough information about both luminance and chrominance components of the noise from the grayscale patches.
Vlad
I have expressed an openness to discussion. That's not an openness to being treated like an idiot.
You can come down from your defensiveness over your target and explain yourself, or you can ignore my request for further discussion (and do a disservice to others). But rational discussion requires that we treat each other as if we are intelligent and striving to come up with a good set of profiles.
(OK. I'm done bristling over your reactions on this thread.)Yes, I do recognize that neutral patches contain chromatic information. Acknowledging that a bright neutral patch has chromatic information does not mean it has the same chromatic information or relative likelihood of generating noise as the saturated cyan patch on the Macbeth ColorChecker card, however.
I do not accept the myth that CCD and CMOS chips are inherently linear devices. Over a normal operating range, they are approximately linear, and a linear model does do a sufficiently good job of mapping input to output. However, the relative likelihood of noise is not equally distributed over all three channels. Nor is it uniform from shadow to highlight. If they were linear, noise would not have a higher likelihood in shadows. Nor would there be a difference between a RAW linear TIFF and a RAW nonlinear TIFF (which attempts to remove nonlinearities in the tails of the operating curve) except for the gamma difference.
If we can agree that CCD and CMOS chips are not perfectly linear and that the relative likelihood of noise is not equal for the separate RGB channels, especially for saturated colors, then it would seem that incorporating color information would likely improve the noise profile.
You've been dismissive without explaining yourself. Educate me and others who care to read this thread. I am willing to admit I am advancing a theory, one which is empirically testable. If you've tested it previously, share the results and save others work. But, if you haven't tested it, you might want to reconsider whether a little more openness of mind might not lead to an improvement in noise profiling cameras like the Canon 10D.
I must tell you, I find you reaction on this thread to be very surprising.
Cheers,
Mitch
Mitch,
First of all, let me apologize if I sounded defensive. I think I misunderstood your intention. I felt the goal was to utilize in the most efficient way the tool which I know how to work with but you probably want to discuss the theory.
Ok, lets do this. I agree that sensors are not linear especially in some ranges. I do agree that noise is not uniform from shadow to light. I am not completely clear about your definition of relative likelihood of noise though.
Yes, color information is important though less important for noise reduction then luminance information. Yes, it can be and is used to make profiles more accurate.
Yes, we have results and have already shared them with everybody - in the form of the user guide with detailed profiling instructions that offer specific calibration target. This is our result - this calibration target provides accurate results and there is no real need to make things more complicated at least at this stage. There are a lot of other problems with digital cameras that can make profiling less accurate besides the target. The calibration method is developed based on extensive experiments made during several years of Neat Image development.
I agree that according to theorists, everything practical is imprecise and bad. But our experience says that the target is good enough for practical purposes. No need to build a more complex theory where a simple one works fine. Satellites still fly according to Newton, not Einstein.
Again, in this forum I try to help people to use NI in the most efficient way. I am not trying to discuss theory here or give out the results of our research work so that competitors had an easier life.
Once again, please excuse me if I gave you not what you expected.
Vlad
First of all, let me apologize if I sounded defensive. I think I misunderstood your intention. I felt the goal was to utilize in the most efficient way the tool which I know how to work with but you probably want to discuss the theory.
Ok, lets do this. I agree that sensors are not linear especially in some ranges. I do agree that noise is not uniform from shadow to light. I am not completely clear about your definition of relative likelihood of noise though.
Yes, color information is important though less important for noise reduction then luminance information. Yes, it can be and is used to make profiles more accurate.
Yes, we have results and have already shared them with everybody - in the form of the user guide with detailed profiling instructions that offer specific calibration target. This is our result - this calibration target provides accurate results and there is no real need to make things more complicated at least at this stage. There are a lot of other problems with digital cameras that can make profiling less accurate besides the target. The calibration method is developed based on extensive experiments made during several years of Neat Image development.
I agree that according to theorists, everything practical is imprecise and bad.
But our experience says that the target is good enough for practical purposes. No need to build a more complex theory where a simple one works fine. Satellites still fly according to Newton, not Einstein. Again, in this forum I try to help people to use NI in the most efficient way. I am not trying to discuss theory here or give out the results of our research work so that competitors had an easier life.
Once again, please excuse me if I gave you not what you expected.
Vlad
I make fine art prints. I put a lot of effort into them. If extra effort in the noise profiles will give me better results, it's an investment of time I'm willing to make.
I certainly do understand the concept of diminishing returns, which I believe is what you're trying to express to me.
I teach statistical research methods and research design at the graduate level. So, when people tell me you'll have to trust me on methodology, I'm uncomfortable.
Your manual is not explicit about your methodology, and you are correct that I would like more explicitness.
You provide an excellent product. One that I've been very comfortable to recommend in multiple forums on the Web. You provide excellent customer service. I think you have nothing to fear by explaining your methodology. In the end, by having many people commenting on the methodology, you may be able to engineer an even better product. It takes more than a methodology or an algorithm to engineer something like Neat Image Pro +.
Cheers,
Mitch
I certainly do understand the concept of diminishing returns, which I believe is what you're trying to express to me.
I teach statistical research methods and research design at the graduate level. So, when people tell me you'll have to trust me on methodology, I'm uncomfortable.
Your manual is not explicit about your methodology, and you are correct that I would like more explicitness.
You provide an excellent product. One that I've been very comfortable to recommend in multiple forums on the Web. You provide excellent customer service. I think you have nothing to fear by explaining your methodology. In the end, by having many people commenting on the methodology, you may be able to engineer an even better product. It takes more than a methodology or an algorithm to engineer something like Neat Image Pro +.
Cheers,
Mitch
Mitch,
I fully understand your desire to perfect your images even if additional efforts and time are required. I just wanted to say that the returns are now going to be noticeable higher if, say, another target was used.
I also understand that you don't trust mere words on methodology. Every scientist wants a convincing proof, I understand. I just don't think this is generally possible or required of a commercial product. - I get Photoshop, I apply, say, its sharpening to a JPEG from my digital camera and there is absolutely no proof this sharpening is done in the right way.
Regarding public discussion on methodology, I think you are right in the sense that in principle this should lead to better methods and final results. Probably even faster than in any other way. This is what science is all about. However, science and market are two different worlds, and it is very difficult to succeed in both at the same time. Say in Astronomy, there are very good (accurate, based on good theory) tools to postprocess data from CCD in telescopes but nobody heard of these tools in the digital camera market.
Anyway, I appeciate your comments and in fact I have written down your idea regarding profiling for the future work. Probably there will be a point when returns will become more noticeable on the background of other factors.
Thank you,
Vlad
I fully understand your desire to perfect your images even if additional efforts and time are required. I just wanted to say that the returns are now going to be noticeable higher if, say, another target was used.
I also understand that you don't trust mere words on methodology. Every scientist wants a convincing proof, I understand. I just don't think this is generally possible or required of a commercial product. - I get Photoshop, I apply, say, its sharpening to a JPEG from my digital camera and there is absolutely no proof this sharpening is done in the right way.
Regarding public discussion on methodology, I think you are right in the sense that in principle this should lead to better methods and final results. Probably even faster than in any other way. This is what science is all about. However, science and market are two different worlds, and it is very difficult to succeed in both at the same time. Say in Astronomy, there are very good (accurate, based on good theory) tools to postprocess data from CCD in telescopes but nobody heard of these tools in the digital camera market.
Anyway, I appeciate your comments and in fact I have written down your idea regarding profiling for the future work. Probably there will be a point when returns will become more noticeable on the background of other factors.
Thank you,
Vlad
Actually, Vlad, Adobe is quite forthcoming about the algorithms in their various tools. For example, the principles behind USM are well-known.
You would not even be able to market Neat Image Pro + if Adobe was not only forthcoming about how their product works but also open in their architecture, providing the API calls and hooks necessary to build add-ins and the documentation in the SDK to take advantage of those features.
It is clear we do not agree on the point of being open about methodology. I am left to speculate that the methodology is actually quite simple. I'm reminded of the Wizard of Oz telling Dorothy to pay no attention to the man behind the curtain.
I do have the feeling that you are considering my comments. That's fair and reasonable.
I'll proceed as I suggested. Nothing you have told me leads me to believe my profiles will fare any worse for the extra effort.
Best wishes,
Mitch
You would not even be able to market Neat Image Pro + if Adobe was not only forthcoming about how their product works but also open in their architecture, providing the API calls and hooks necessary to build add-ins and the documentation in the SDK to take advantage of those features.
It is clear we do not agree on the point of being open about methodology. I am left to speculate that the methodology is actually quite simple. I'm reminded of the Wizard of Oz telling Dorothy to pay no attention to the man behind the curtain.
I do have the feeling that you are considering my comments. That's fair and reasonable.
I'll proceed as I suggested. Nothing you have told me leads me to believe my profiles will fare any worse for the extra effort.
Best wishes,
Mitch
All of this WAS open before Photoshop v7.0. Not anymore, unfortunately. I think the main reason why they made the SDK non-public is the market pressure.You would not even be able to market Neat Image Pro + if Adobe was not only forthcoming about how their product works but also open in their architecture, providing the API calls and hooks necessary to build add-ins and the documentation in the SDK to take advantage of those features.
Vlad
The SDK is still available for PS CS. You just have to subscribe to their developer program. $195 a year. It also includes working copies of products.
I don't mean to imply you're being greedy or anything of that sort, Vlad. I think your product is very reasonably priced. You're especially generous about upgrades. The upgrade from Pro to Pro + was only $10 and the upgrade was free. One could not ask for more reasonableness in how you sell Neat Image.
Cheers,
Mitch
I don't mean to imply you're being greedy or anything of that sort, Vlad. I think your product is very reasonably priced. You're especially generous about upgrades. The upgrade from Pro to Pro + was only $10 and the upgrade was free. One could not ask for more reasonableness in how you sell Neat Image.
Cheers,
Mitch
No products with $195 version. Anyway, the point is not price - I wouldn't hezitate to pay if the result was a matter of days (it is months now) and was guaranteed (it is not). Besides, filling out 8+ pages of application forms is overkill too. I understand they fight with other big companies that make plug-in-compatible products. Who suffers the most are the small independent developers.gmitchel wrote:The SDK is still available for PS CS. You just have to subscribe to their developer program. $195 a year. It also includes working copies of products.
Vlad
P.S. Funny thing, they don't have the country of my current residence in their country list.