Resolution, sharpness and colordepth.

[h-munster]

Resolution is mostly independent of sharpness.

As I mentioned earlier, sharpness and resolution are not the same thing. However, the two qualities are not independent. Resolution affects sharpness -- if the resolution is too low, even high contrast edge details look blocky and soft.

Yes, a high resolution image can look sharp when scaled down, but an image that is native 320x240 can still look sharp to our eyes. One person could say a picture looks too sharp, and another can say the same picture looks just ok. Its all relative.

Let's not forget that screen size and viewing distance are also significant.

 

[h-munster]

However, if the digitizer does not have the sensitivity to capture the variation, the resolution will have no effect on the resulting image.

Not sure what is meant by "sensitivity to capture." Does it mean color depth? ...bit depth? ... the dynamic range of the capture? Each of these three qualities are different (but not necessarily independent of each other).

An example would be to capture an image that has an 8 bit digitizer, then representing that as a histogram, using 16 bit values. One will have a histogram of 'spikes' corresponding to the 8 bit values, and 'zero' in between. Then compare that to an image digitized to 16 bits, and one will see a histogram with values lying between the '8 bit' samples.

Okay. "Sensitivity to capture" evidently means "bit depth."

Although this example is correct, we were not trying to jump back up to a "higher quality" stage in our scenario -- we were just going from 35mm and 65mm onto DVD. The final 16-bit histogram merely demonstrates that we don't have 16-bit bit depth.

The actual color depth (and the number of "in-between" tones) is not always represented by a histogram. Even with the 8-bit bit depth and the spiked, 16-bit histogram, we can easily retain the same color depth as that of a standard HD with 16-bits of bit depth! Just a little higher resolution in the digitizer will maintain such color depth, in spite of the lower bit-depth and the spiked histogram.

To put it another way, lack of bit depth will appear as spikes in a histogram, but higher color depth (due to higher resolution) will produce the "in-between" tones that are not represented in the "zero" troughs of the histogram.

There are more details on how color depth differs from bit depth on this page.

 

[h-munster]

With any fixed pattern digital sensor system, Bayer or not, your visible resolution will be at the very best around 85% of the pixel resolution unless you are willing to tolerate aliasing and moire in the finished image.

In regards to aliasing and moire', a lot depends on the resolution of the sensor. Higher resolutions tend to have less of a problem. It is doubtful that one would be forced to use only 85% of an 25-Megapixel, RGB sensor, to avoid moire'/aliasing.

Also with digital systems, oversampling, ie higher resolution, greater color depth. greater dynamic range in the original will result in better image quality at finish resolution. You need 2x oversample to fully resolve to your finished format and differences will be noticeable up to 4x oversample. On TV for instance the difference in an SD program that originated in HD vs SD that was shot in SD is very obvious.

This assertion is one side of the primary topic of "contention" of this discussion. What constitutes "better image quality" is in question, as is how and when the image differences become perceptible.

 

[j1clark@ucsd.edu]

The actual color depth (and the number of "in-between" tones) is not always represented by a histogram. Even with the 8-bit bit depth and the spiked, 16-bit histogram, we can easily retain the same color depth as that of a standard HD with 16-bits of bit depth! Just a little higher resolution in the digitizer will maintain such color depth, in spite of the lower bit-depth and the spiked histogram.

I read the article and I think I understand. However, having the triple, RGB, and various values for each, do not, relative to human perception, result in a 30 bit 'color space' for a 10 bit digitization of R,G, and B, that humans can uniquely perceive.

What this sort of thing may suggest is the 'proper' viewing distance for a given resolution, given the human perception.

 

[Patryk_Rebisz]

wow, you guys are geeks...

 

[Joe Calabrese]

These images are low in resolution, but are still relatively sharp (especially the bottom left one).

 

[h-munster]

I read the article and I think I understand. However, having the triple, RGB, and various values for each, do not, relative to human perception, result in a 30 bit 'color space' for a 10 bit digitization of R,G, and B, that humans can uniquely perceive. What this sort of thing may suggest is the 'proper' viewing distance for a given resolution, given the human perception.

Technically, the bit depth of each pixel and of each RGB cell is straightforward and simple, and the bit depth doesn't vary, regardless of resolution, viewing distance, magnification, etc.

However, a change in the viewing distance, magnification or resolution will definitely affect the color depth.

 

[h-munster]

These images are low in resolution, but are still relatively sharp (especially the bottom left one).

Thank you for posting those images.

I have enlarged a section of the bottom left photo which included the distant mountain tops. Notice that we now see the individual RGB pixel cells:

This is the same photo, unaltered. It is just magnified, which gives a lower resolution -- the original image pixels are now humongous.

Does this lower resolution look sharp?

Try squinting or back off at least two meters from the monitor. Notice how the image looks sharper? That's because, by squinting or by backing off, you are blending the pixels, which is the same thing that happens when the resolution increases. Also, by increasing the viewing distance, you are actually increasing the resolution (per degree of view angle).

In addition, please note that this magnified version shows streaks and "banding" in the sky area. In the original, higher resolution version of this image which you posted above, these bands blend to form a smooth gradation, which demonstrates the increased color depth that accompanies higher resolution.

It is a fundamental photographic principle that increasing resolution boosts the color depth and sharpness. However, bit depth additionally affects color depth just as the contrast of the edge details additionally affects sharpness.

 

[h-munster]

wow, you guys are geeks...

Dude! The upshot of this thread is that your Red footage isn't much better than recording the HDMI output from an HV20, but the extra resolution of your Red actually gives richer colors.

 

[Otis Grapsas]

Interesting thread.

The dynamic range is inside a single pixel. It's saturation vs noise. I'm taking about the spec, not how your highlights or shadows look, the apparent dynamic range, how the camera handles overexposure and anything like that, is an issue of design, the sensor is a very small part of it. The right gamma curve of film emulation will look superior using the same hardware. It's 50% of the cinematic look. Lenses are important. The sensor is also very important. The rest of the cinematic look is not equipment, it's knowledge and talent and no white paper adresses that. Resolution is not very important if you can capture it at low compression.

Resolution does not change color depth. That article on color depth is simply nonsense. Color does not vary fast enough in nature to make topical summing of higher resolution important.

The mathematical gain in S/N is also not that important compared to the other factors.

Resolution will not improve tonality near clipping, because clipping appears over large areas, not single pixels, summing more than a few clipped pixels will only give you clipping, not an intermediate tone. It will marginally soften the edges near the clipped area, but that's hardly an improvement.

The true color depth of the system depends on some factors:

- How many photons does the system capture per pixel? A very small pixel with a compromised quantum efficiency could capture so few photons that its color depth would be severaly compromised.

- What is the noise per pixel? Dynamic range below noise is not very dependable.

-Is there pattern noise and how is it treated? Critical in CMOS sensors.

- How many bits are used for capturing the image (AD converter)? Higher is obviously better.

- What are the required digital gains for white balance? On a typical sensor you could be wasting 1 full bit on Red and Blue to balance a powerful green. On tungsten it could be 2 in blue and none in Red. Optical white balancing saves these losses. The optical white balance filters in cameras are rarely designed for full analog balancing. The engineers do not want to waste the power green that provides extra sensitivity.

- Is there analog or digital gain used? Both have advantages and disadvantages. A digital 2xASA in post compromises each color channel by 1 full bit, your 18% grey quality contains half the tonality in each color down to the noise floor, and 1/8th the total tonality (-3 bits). That's because you move your 18% grey 1 stop lower on the digital exposure through the gamma curve, it's not true sensitivity. The equivalent 6dB of analog gain maintains the bit precision because the 18% grey hits the same point on the AD converter, unlike digital gain, it happens before the AD. Digital 2x ASA increases aliasing but it allows an extra stop in highlight range if there is provision for shaping the highlights with a gamma curve so they are not lost. Analog gain is free from aliasing but wastes one stop. With enough AD precision, digital gain will work better. It depends on the usefulness of the extra overexposure range. Use the minimum gain setting and digital ASA setting to get the best results.

- Is there compression involved? Compression reduces color depth, compressed color suffers in low data rates, it's artifacted in depth and space. Even when the image looks sharp and detailed, your color is forever gone. It will have that digital look. There will be highly saturated tones but there will be nothing between the tones.

- If its bayer, is the algorithm good? Most debayer algorithms are designed for high luma sharpness, not for high color quality.

- How heavy is the color matrix? The color matrix is mixing color in positive and negative amounts to transform the sensor color space to the desired viewing color space. That involves math on your color values. If the bayer sensor filter or 3CCD color filters are good, the output will not require a color matrix with heavy parameters, so color quality will be better. A small pixel CMOS sensor will be very far from ITU709 in saturation and color primaries, that it will produce inferior results.

- Is the UV/IR filter good? On most camcorders it's nothing special. Usually it's compromised to get as much IR in as possible in order to gain sensitivity in tungsten lighting. A weak IR filter will mess with your tonality. It will inject grey and IR in image of many fabrics and natural objects and reduce color purity. They could have additional grey or magenta. The color matrix cannot fix it, the compromise will still be very important. Install an additional IR to purify your color.

- Exposure is critical of course. Underexposure compromised precision which compromises tonality.

- Are the gamma curves good? A mediocre gamma curve does not produce very good saturation, it requires post processing which is usually performed on compromised compressed material.

- Is the post processing software designed for high precision? Float or 16bit integer at least?

- How is the bit precision of your files? Processing low bit depths will compromise tonality, especially if the image does not look that good out of the camera. 8bit uncompressed is quite good, 10bit log or 12bit linear RGB are excellent. 8bit YCbCr is mediocre, 10bit YCbCr is good. RGB is about 2 bits better than YCbCr at the same precision (Think 24bit vs 22bits). If the camera does not output something better than 8bit, convert to something better if you intend to do multiple passes of processing. Multiple passes on 8bit is a very serious mistake, if it's compressed, it's a crime. If you have a specific goal such as ITU709 and you have an option in camera, try to get a finished look in camera, before bit reduction and compression is applied on your image. GE the gamma and exposure as close as you can to the final look. Never use flat settings on low bit depth and high compression cameras. You get the range but the output will be aliased.

- If you can afford it, keep everything uncompressed from the sensor to the finishing and avoid YCbCr downsampling and workflow codecs. Uncompressed recording, proxy editing, uncompressed finishing and creation of delivery format from the high bit depth uncompressed material.

If you do all this you might be able to get a 21 to 22bit of tonality out of a 12bit uncompressed sensor after the processing. Some high end medium format backs with large pixels will give you up to 25 but that's it. The tonality will transfer well to a bluray assuming a high quality encoder.

 

[Otis Grapsas]

I read the full "color depth" "article". I have never read anything like it!

The basic problem with the thought process on that "article" is that it applies the multiplication we use in RGB combination (which are independent colors so 1024x1024x1024=30bit) to the addition required when we downsample/bin pixels of the same color (1024+1024+1024+1024=12bit).

When we are downsampling or binning we are adding, not multiplying. So, a 4K sensor downsampled to 2K will not produce stratospheric color depth performance. It's just a 12 bit (0 to 4095) value vs four 10 bit (0 to 1023) values and it will certainly not store 38 bits! The best case scenario would be matching the color depth performance of a 2K Sensor of the same sensor size. What high resolution does provide is higher luma and chroma resolution/MTF after downsampling.

"So, we have a 51,580,835,258,454,554,764,640,256.25-times expansion in color depth! This unfathomable increase in color depth is produced by a mere 4-times resolution increase."

This reminded me of Ali-G presenting his ice-cream gloves (3:10):

http://www.youtube.com/watch?v=PDu9CvbrnlM

 

[h-munster]

The dynamic range is inside a single pixel. It's saturation vs noise.

Not sure why everyone keeps bringing up dynamic range. The topic is color depth, resolution and sharpness and how the three are interrelated.

Dynamic range and color depth are two very different, independent properties. One could have 200db of dynamic range, but, if there are only 8-bits per color channel, each RGB cell will only be able to produce of 16,777,216 colors, nevertheless. So, the color depth of each RGB cell is independent of the dynamic range.

I guess that it would be wise to also clarify what is meant by "saturation" in the context of dynamic range -- saturation of a pixel is it's maximum quantifiable input level. I only mention this so that someone reading this thread does not confuse dynamic range saturation with "color saturation."

I'm taking about the spec, not how your highlights or shadows look, the apparent dynamic range, how the camera handles overexposure and anything like that, is an issue of design, the sensor is a very small part of it.

Not sure what this sentence means.

The right gamma curve of film emulation will look superior using the same hardware. It's 50% of the cinematic look. Lenses are important. The sensor is also very important. The rest of the cinematic look is not equipment, it's knowledge and talent and no white paper adresses that.

Not that this subject is "on-topic", but I would say that the two most important "film-look" characteristics are the frame rate (24 fps) and the scanning method (progressive). Both properties are taken for granted these days, as most cameras have these capabilities built-in.

Gamma is usually not that important in simulating film, unless one tries to use it to match the washed-out look of a film negative when it first hits the telecine. However, probably very few bother with that technique these days.

There are "video-specific" problems that can reveal a non-film image, such as the peculiar look of blown-out highlights, and, of course, interlaced scan with no pulldown. I guess the blown-out highlights could be considered a result of a certain gamma, but I have always used the term to refer to the general "slope" of "curve" of the tones between the extreme highlights and shadows.

Shallow DOF, lenses, lighting, camera movement, etc. -- none of these really contribute to a film look. Plenty of movies shot on film have had deep DOF, bad lenses, bad lighting and no camera movement, and all of them nevertheless looked like film, not video. Likewise, plenty of videos have had shallow DOF, great lenses, superb lighting and professional camera movement, and all of those videos nevertheless look like video.

Furthermore, if you take film footage and speed it up above 24fps, it looks like video!

Really, film look is mostly in the frame rate and the scanning method.

Resolution is not very important if you can capture it at low compression.

Well, the importance of resolution is one of the questions of this thread. I happen to think that 1920x1080 is enough resolution for 98% of cinematography. I know plenty of folks who can produce compelling images with a Pixel-Vision camera.

Resolution does not change color depth. That article on color depth is simply nonsense.

Actually, resolution has a very strong influence on color depth. This principle is fundamental to every photographic medium, in both the capture and display of images.

Higher resolution begets greater color depth, and this reality was demonstrated clearly with the low res sky banding, posted earlier in this thread. There are plenty of other obvious examples of this fact. I would be happy to supply another.

Color does not vary fast enough in nature to make topical summing of higher resolution important.

Not sure what this statement means.

The mathematical gain in S/N is also not that important compared to the other factors.

Okay, but it's not really related to the topic of this thread.

Resolution will not improve tonality near clipping, because clipping appears over large areas, not single pixels, summing more than a few clipped pixels will only give you clipping, not an intermediate tone.

That is true if the illumination level is above the clipping threshold. When it is just below the clipping threshold, then the slight, extra tonality of high resolution comes into play.

It will marginally soften the edges near the clipped area,...

If by "soften", you mean "add another shade or two," then that is what I have been saying.

... but that's hardly an improvement.

Agreed. It's not very substantial, compared to other methods which increase dynamic range

The true color depth of the system depends on some factors:...

The long, detailed list of variables that follows is both informative and impressive. However, most of the items on the list do not affect color depth as fundamentally as resolution. Resolution's effect on color depth transcends all of those variables -- higher resolution will always yield greater color depth, regardless of the given status of any of the properties in the list.

Most of the items on the list are "dependent variables" which rely on other contingencies to affect color depth. For instance, consider dynamic range, gain, quantum efficiency, gamma, etc. Expansion of such properties will not increase the color depth, unless the bit depth in the AD conversion section increases. They are dependent variables, whose effect on color depth is less transcendent than that of resolution.

Furthermore, resolution transcends the listed variables in another important way -- its effect on color depth also applies to film emulsion and to all other photographic media. All of the variables on the list (except for one) are limited only to video/digital imaging. Some of the listed items pertain only to digital capture and even more specifically, only to CMOS sensors and Bayer sensors.

The one exception on the list that applies to non-video/digital is the point relating to exposure. However, suggesting that exposure affects color depth is like saying that shooting with the lens cap mounted affects color depth. Changing the exposure doesn't really affect the color depth capability of the system; setting exposure merely allows certain scene tones/colors to be captured (or not captured) by the system.

 

[Otis Grapsas]

Color depth is interconnected to dynamic range. The dynamic range of the three colors, affected by the parameters I listed, produces the color depth. It's the different shades a sensor can capture.

There are no 200dB dynamic range systems that use 8bit converters. The converters are always better than the sensor. Many 10stop cameras come with 12/14bit converters because they want to capture the noise. In high quality designs nothing less than 12bit is acceptable today because of the requirements I mentioned (digital while balance, gamma application, matrix math etc).

A gamma curve is what produces most of the film look IMHO. All the characteristics we love, the highlight handling, the high contrast mids, the velvet blacks.

1920x1080 is enough, but only if you capture without introducing artifacts as large as 8x8 pixel blocks. An intraframe codec will do that in static shots, a GOP codec will also do it in motion and HD broadcast will do the same. It will never be as good as uncompressed even though it has the same resolution. The artifacts are far larger. Up to 64x in area for mpeg2.

That article is mathematically flawed and those who reach similar conclusions should check their math. As I explained, using 4x smaller pixels in order to get 4x megapixels does not improve color depth vs the original larger pixel. It compromises dynamic range and color depth, both by 2bits per channel. If the downsampling gives that back in "color depth", the 2 bits in dynamic range are lost forever. You have an inferior result and you exchanged dynamic range for higher resolution. Everything comes at a cost.

Most of the parameters I listed also apply in film, both optical and digital because film is processed digitally and delivered digitally. Photons are captured by film, film has a grain level, bit depth is important in DI, optical white balance is critical in film, digital white balance is commonly applied, digital gain is applied any time you compensate exposure in film or change the look of a DI image, all the post processing quality parameters are as valid as on a digital camera, etc.

 

[Otis Grapsas]

Film is quite similar to digital today. The entire chain stays the same. The only difference is that film exposure and negative processing is added on this chain before the sensor.

Digital: Light->optical filter->lens->sensor->AD->digital processing and output

DI Film: Light->optical filter->lens->analog film exposure and processing->sensor->AD->digital processing and output

 

[h-munster]

II read the full "color depth" "article". I have never read anything like it!

Yes. Nobody ever mentions how resolution affects color depth.

The basic problem with the thought process on that "article" is that it applies the multiplication we use in RGB combination (which are independent colors so 1024x1024x1024=30bit) to the addition required when we downsample/bin pixels of the same color (1024+1024+1024+1024=12bit). When we are downsampling or binning we are adding, not multiplying. So, a 4K sensor downsampled to 2K will not produce stratospheric color depth performance.

There is no "downsampling" involved. The pixels/RGB cells interblend in the viewers eye at the point of display.

Three adjacent pixels (red, blue and green) interblend to form the single color of the RGB cell. When we have multiple RGB cells in the same space that we formerly had a single RGB cell, the multiple RGB cells interblend to form a single color -- just like the red, blue and green pixels interblended in the single RGB cell.

Thus, the interblending math for the multiple RGB cells is identical to the math for the multiple RGB cells, with one catch. We multiply in both instances, but with the multiple RGB cells, we have to allow for a "catch" -- some permutations of multiple RGB cells yield several versions of the same combination of colors. So, the formula for multiple RGB cells has to "weed-out" all of these redundant combinations.

"So, we have a 51,580,835,258,454,554,764,640,256.25-times expansion in color depth! This unfathomable increase in color depth is produced by a mere 4-times resolution increase."
This reminded me of Ali-G presenting his ice-cream gloves (3:10): http://www.youtube.com/watch?v=PDu9CvbrnlM

That's a funny joke, but that figure is true.

There are reasons why we don't see a dramatic color depth change when we view standard HD and hyperHD side-by-side.

First of all, most standard HD is 10 bit (per color channel), which means that we already start out with 1,073,741,824 possible colors per RGB cell (that's a lot).

Secondly, just because we have all those possible colors, it doesn't mean that they will all be utilized (a white wall will look like a white wall, regardless of the color depth).

Lastly, unless we use some sort of threshold/sensitivity map on the individual RGB cells, the extra color depth from higher resolution will mostly be captured in the more intricately patterned areas of the image. Perhaps this circumstance is what you were getting at in an earlier post when you said, "Color does not vary fast enough in nature to make topical summing of higher resolution important." By the way, there are currently a couple of cameras that use sensitivity mapping.

 

[Otis Grapsas]

They don't mention resolution because nobody would think to increase resolution to improve tonality. It's a bad idea for the practical reasons I mentioned.

RGB are independent. You know what that means mathematically? You can't multiply to find the combination. 4 grey pixels are not independent, you have to add. They wouldn't be indepedent if you were able to see all of them one by one, and that would be resolution, not tonality. That article is laughable. The math is laughable. The result is a joke. Anyone with common sense would stop and think after that number appeared on the calculator screen.

You want to rule out all unique combinations? All 4 value combinations that add to the same result should be ruled out. That's just one combination perceptually. That's it. You will get 2bits for 2x reduction. That happens on a 2x2 filter and on any filter of similar power. Photons add, in pixels, in their digital signal represenation, everwhere.

RGB combine in our vision because our color resolution is very low. This doesn't happen in this case.

If you accept that they combine you have to find how they combine and present a mathematical representation of the optical filter in our eyes. There are available models for that. An MTF curve, a mathematical filter. If you can still see half luma on the output, the mathematical representation cannot be better than 2 pixels wide equivalent. 4 into 1. That's about 2bits for all available filter types that still allow to discern one 4pixel block from the next.

The standard HD is 10 bit with LOG or GAMMA encoding. It's far greater precision in the ADs because these sample linearly. We need the extra precision.

I have seen 2x reductions vs the original in resolution from a range of distances and I never noticed any improvement in color depth. It's not there. Try it. Do you know of anyone that moved his chair to twice the distance from the screen and discovered a color depth improvement of this magnitude? At best, he would discover the compression artifacts are harder to see or that broadcast HD is less annoying.

I just want to know who started this joke

 

[h-munster]

Color depth is interconnected to dynamic range. The dynamic range of the three colors, affected by the parameters I listed, produces the color depth. It's the different shades a sensor can capture.

I agree that the number of colors that an RGB cell can capture is the color depth of that cell. I agree that a reduced capture dynamic range can effectively limit the number of colors that an RGB cell can capture -- in that sense dynamic range and color depth are interconnected.

However, those parameters in the list mostly involve what is currently practical in regards to sensors and signal processing technology. They are not overiding photographic principles, which would also apply to film emulsion, halftone/screen printing, monitors, photographic prints, etc.

Practical limits of digital imaging are constantly being pushed outward, and some of the limits posed by those parameters will become less relevant. Yet, the relationship between color depth and resolution will always exist, regardless of new limits and of the type of photographic medium.

There are no 200dB dynamic range systems that use 8bit converters.

I was using an unlikely example to demonstrate that bit depth is different from dynamic range.

A gamma curve is what produces most of the film look IMHO. All the characteristics we love, the highlight handling, the high contrast mids, the velvet blacks.

Take a VHS tape of a movie shot on film (with its inherent gamma) and "fast forward" it. Then, see if the fast forwarding footage doesn't suddenly look like video.

1920x1080 is enough, but only if you capture without introducing artifacts as large as 8x8 pixel blocks. An intraframe codec will do that in static shots, a GOP codec will also do it in motion and HD broadcast will do the same. It will never be as good as uncompressed even though it has the same resolution. The artifacts are far larger. Up to 64x in area for mpeg2.

I would say that it needs to be completely uncompressed for theatrical exhibition. Any artifacts in the movie theaters are difficult to stomach.

That article is mathematically flawed and those who reach similar conclusions should check their math.

To find the color depth of a group of RGB cells, we multiply.

As I explained, using 4x smaller pixels in order to get 4x megapixels does not improve color depth vs the original larger pixel. It compromises dynamic range and color depth, both by 2bits per channel.

Again, you are referring to practical limitations of certain sensor technology. Does this compromise come into play in the same way with quantum dot sensors? ... or with film emulsion?

Furthermore, what kind of dynamic range limits would we encounter if we start out with a 40x40-microns pixel, which would mean that the 1/4x smaller pixel is 20x20-microns?

If the downsampling gives that back in "color depth", the 2 bits in dynamic range are lost forever. You have an inferior result and you exchanged dynamic range for higher resolution. Everything comes at a cost.

There is no downsampling involved.

Most of the parameters I listed also apply in film, both optical and digital because film is processed digitally and delivered digitally.

Sometimes, but not always. Anyway, when I refer to film, I mean emulsion only.

Photons are captured by film, film has a grain level,...

Yes, and the finer the grain, the greater the color depth (all other variables remaining the same).

... bit depth is important in DI,...

Agreed.

optical white balance is critical in film, digital white balance is commonly applied,...

I would say that "optical white balance" is more important in digital. I can over-expose film 3/4-stop and get back a lot of the blue layer while still maintaining nice highlights.

digital gain is applied any time you compensate exposure in film or change the look of a DI image,...

Not sure what is meant by this statement.

... all the post processing quality parameters are as valid as on a digital camera, etc.

Okay.

 

[Otis Grapsas]

I noticed that you don't understand the difference between combining different things (red, green, blue) and same things (monochromatic light, photons, signal produced by counting photons on a sensor). This is key to understanding why multiplication is the wrong thing to do and how flawed that article is. In mathematics it's easy to get lost in numbers. You can start with something simple to understand how it works, then try to find the generic solution. That's how most things are put into an equation. Even natural laws. Equations are not written on a cloud, and they are not in a book until someone puts them in there.

Check out these pairs of RGB values in an 1 bit system:

1. (0,0,0)
2. (0,1,0)
3. (0,0,1)
4. (0,1,1)
5. (1,0,0)
6. (1,1,0)
7. (1,0,1)
8. (1,1,1)

All are unique colors. 2x2x2=8 combination. We multiply. A red pixel is obviously different from a blue pixel and a green pixel even though they have the same power. The filter on top of them makes the difference, they are counting (sensor) or coloring (LCD display) different parts of the spectrum.

And some 3 pixel pairs containing the same thing (monochromatic color) which are filtered optically or digitally to produce the tonal perception:

1. (0,0,0) black=0
2. (0,1,0) dark grey=1
3. (0,0,1) dark grey=1
4. (0,1,1) light grey=2
5. (1,0,0) dark grey=1
6. (1,1,0) light grey=2
7. (1,0,1) light grey=2
8. (1,1,1) white=3

Do you agree with the above? That all light greys will apear equal?

YES! If you agree you are accepting that you can't see detail inside the 3 pixel area and that summing is the right thing to do.

NO! If you don't agree, you are accepting that you can see detail inside the block. This means you can see higher resolution and not tonality. If you can tell no2 from no3 and no5, then congratulations, you obviously have enough resolution in your vision to resolve 3x the horizontal resolution and that's what the sensor will give you.

So for 1 bit:

1 pixel combinations 0,1 = 1*2^1 - (1-1)
2 pixel 0,1,2 = 2*2^1 - (2-1)
3 pixel 0,1,2,3 = 3*2^1 - (3-1)
4 pixels 0,1,2,3,4 = 4*2^1 - (4-1)

And for 8 bits:

1 pixel 0,1...255 = 1*2^8 - (1-1)
2 pixels 0,1...510 = 2*2^8 - (2-1)
3 pixels 0,1...765 = 3*2^8 - (3-1)
4 pixels 0,1...1020 = 4*2^8 - (4-1)

So, for k pixels with N bits of precision, it's k*2^N - (k-1).

For high N, k-1 is worth nothing, so it's practically k*OriginalCountingPrecision, 4x for 4 pixels which is two bits.

 

[h-munster]

I noticed that you don't understand the difference between combining different things (red, green, blue) and same things (monochromatic light, photons, signal produced by counting photons on a sensor). This is key to understanding why multiplication is the wrong thing to do and how flawed that article is.

Let us wait and see below how flawed is the article and the math.

In mathematics it's easy to get lost in numbers.

It certainly is!

Check out these pairs of RGB values in an 1 bit system:

1. (0,0,0)
2. (0,1,0)
3. (0,0,1)
4. (0,1,1)
5. (1,0,0)
6. (1,1,0)
7. (1,0,1)
8. (1,1,1)

All are unique colors. 2x2x2=8 combination. We multiply. A red pixel is obviously different from a blue pixel and a green pixel even though they have the same power. The filter on top of them makes the difference, they are counting (sensor) or coloring (LCD display) different parts of the spectrum.

And some 3 pixel pairs containing the same thing (monochromatic color) which are filtered optically or digitally to produce the tonal perception:

1. (0,0,0) black=0
2. (0,1,0) dark grey=1
3. (0,0,1) dark grey=1
4. (0,1,1) light grey=2
5. (1,0,0) dark grey=1
6. (1,1,0) light grey=2
7. (1,0,1) light grey=2
8. (1,1,1) white=3

Do you agree with the above? That all light greys will apear equal?

Yes

YES! If you agree you are accepting that you can't see detail inside the 3 pixel area and that summing is the right thing to do.

No.

You neglected to try these scenarios with the formula from the resolution/color depth page, didn't you?

As I already mentioned above in post #35, "We multiply... , but the formula... has to 'weed-out' all of these redundant combinations." And weeding-out the redundant combinations in your one-bit monochrome scenario is precisely what the color depth formula will do.

Here is the resolution/color depth formula:
(R+N-1)! / (R!x(N-1)!) = C

"R" is resolution -- the number of RGB cells (or non-unique pixels) in a given group; "N" is the number of available colors/shades of each RGB cell (or non-unique pixel); and "C" is the resulting color depth.

Let's try this formula with your one-bit RGB example.

There is one RGB cell (with three unique pixels), so, R = 1.

There are three, unique, one-bit pixels in each RGB cell, thus, N = 2^1 x 2^1 x 2^1 = 2^3 = 8

We enter these figures into the color depth formula:
(1+8-1)! / (1!x(8-1)!)
= (8x7x6x5x4x3x2) / (7x6x5x4x3x2)
= 8

So, according to the color depth/resolution formula, there are 8 colors possible from your one-bit-per-channel, RGB. How does that compare with your straight multiplication? Oh! It is the exact same result!


Now, let's try the color depth formula with the three, one-bit, monochrome pixels.

There are three non-unique pixels, so, R = 3.

Each non-unique pixel has 2 shades each, so, N = 2

Entering the figures into the color depth formula, we get:
(3+2-1)! / (3!x(2-1)!)
= (4x3x2) / (3x2)
= 4

The color depth formula says that your group of three, one-bit, monochrome pixels should generate four unique shades. How many shades did you get from your manual "summing" method? Oh! You got 4 shades, too!!

By the way, have you noticed anything peculiar about this scenario? How is it that we can get 4 distinct shades from one-bit pixels that only produce 2 shades? Somehow, in your example, the color depth (or luminance depth) is greater than just one-bit. It is because the pixels are small enough so that three pixels interblend to create a single shade -- they have a high enough resolution. Your own example has just proven that higher resolution raises the color/luminance depth!

So far, the math that you have criticized is producing results that perfectly match your own manual efforts, and the article that you have belittled is ringing true in regards to resolution's influence on color depth.

The color depth formula is very precise and it automatically eliminates the redundant permutations. Furthermore, by your own example, increasing the resolution boosts color depth.

What was it that you were saying about me not understanding combinations and about how easy it is to get lost in mathematics?

By the way, in any scenario that has a group of one-bit, monochrome pixels interblended, the number of possible shades from the group will always be the number of pixels in group plus one. For instance, your example has a group of 3 one-bit, monochrome pixels, therefore, the number of shades possible from this group will be 4. If you had a group of 10 one-bit, monochrome pixels, they would interblend to form 11 shades. The color depth formula bears this out, as will your manual method.

However, your manual summing method becomes impractical when we deal with groups of RGB cells with higher bit rates. There is an example of demonstrating the color depth of a group of four, interblended RGB cells in the resolution/color depth page.

NO! If you don't agree, you are accepting that you can see detail inside the block. This means you can see higher resolution and not tonality. If you can tell no2 from no3 and no5, then congratulations, you obviously have enough resolution in your vision to resolve 3x the horizontal resolution and that's what the sensor will give you.

So for 1 bit:

1 pixel combinations 0,1 = 1*2^1 - (1-1)
2 pixel 0,1,2 = 2*2^1 - (2-1)
3 pixel 0,1,2,3 = 3*2^1 - (3-1)
4 pixels 0,1,2,3,4 = 4*2^1 - (4-1)

And for 8 bits:

1 pixel 0,1...255 = 1*2^8 - (1-1)
2 pixels 0,1...510 = 2*2^8 - (2-1)
3 pixels 0,1...765 = 3*2^8 - (3-1)
4 pixels 0,1...1020 = 4*2^8 - (4-1)

So, for k pixels with N bits of precision, it's k*2^N - (k-1).

For high N, k-1 is worth nothing, so it's practically k*OriginalCountingPrecision, 4x for 4 pixels which is two bits.

Not sure how relevant this is, now that the soundness of the color depth formula has been demonstrated.

Any way, when posting a formula, please explain it.

 

[h-munster]

They don't mention resolution because nobody would think to increase resolution to improve tonality. It's a bad idea for the practical reasons I mentioned.

No one has suggested increasing resolution to gain color depth. I was merely stating the often neglected fact that higher resolution boosts color depth.

I have seen 2x reductions vs the original in resolution from a range of distances and I never noticed any improvement in color depth. It's not there. Try it. Do you know of anyone that moved his chair to twice the distance from the screen and discovered a color depth improvement of this magnitude?

The rise in color depth that accompanies an increase in viewing distance was demonstrated earlier in this thread. The banding in the sky of the enlarged section of the photo turns into a smooth gradation (blends) as one moves away from the screen.

Increasing the viewing distance heightens the angular resolution of the image being viewed (the pixels get smaller). Hence, adjacent RGB cells eventually start to interblend, and the color depth increases.