Toothwalker
Offline
• • •
Upload & Sell: Off
|
 p.3 #4 · Zeiss Thread Image Processing/Shooting Techniques | |
Samuli Vahonen wrote:
All of this comes to basics of processing images "naturally" or "unnaturally". In "nature" light and these things are linear. In computer due to using CRT tubes in past we have ended up to using gamma to display images correctly. However all formulas in world are linear unless they have been "compensated"/"corrected" to take the gamma into account. However in graphics almost all software does the tricks without this "compensation"/"correction" while most of the people only process only gamma raped images. Weird, but many things man have created are weird and not very logical... BTW. Also if you work with tools like dcraw, you notice that our camera sensors also output the data in linear (=gamma 1.0) mode.
...Show more →
When I say that I don't do color management I mean that I do not actively bother with it and just use all available tools with their default settings. At one point I started to look into these matters, and got an incredible headache as well as a system crash. For instance, if you say that all formulas in the world are linear, I wonder "linear in what?" You may answer "Linear in the pixel value if the specified gamma value is one," but that does not really answer my question as I want to know what physical or other quantity these pixel values represent. Should the input to the formula be linear in the amplitude, or linear in the intensity of the light wave incident on the sensor? Is the sensor response linear, and what about the RAW converter? Usually the downscaling is the last step in a process involving adjustment of levels, contrast, curves, and whatever. What do these operations do to the linearity of whatever goes into the formula? To me, the gamma factor is only one of many unknown distortions.
Anyway, my whole color management consists of opening a Canon RAW file with LR4 and outputting an sRGB TIFF file. DPP offers the option to embed an ICC profile, but I can't find this option in LR4. Perhaps it is included by default, but then it is lost again after I do some bookkeeping tasks with Matlab. My apologies for that.
I asked color managed and 16-bit in order to process images in controlled manner with "civilized" and commonly used tools (e.g. PhotoShop). However 16-bit source image is not color managed so I have to use very basic and very hard to use tools (for common population). Commonly used tools (e.g. PhotoShop) don't have good and accurate ways to adjust gamma unless you do it via color management (well considering the level of average software user that is just a good thing...). Having image in 16-bit allows one to do translations between ICC-profiles (or different gammas) without losing image quality, this is usually causing color and other artifacts and data loss when done in 8-bit.
Due to not having color managed images, all processing is done in ImageMagick command line tools - it's free tool and can be used in all common platforms (and converted from source code to less common platforms): http://www.imagemagick.org
If we look at ring chart and the differences to it when lanczos scaling is done to it:
1. Gamma 1.0 (or thereabouts, I assumed you had your non-color managed image in gamma 2.2)
2. Gamma 2.2 (or thereabouts)
...Show more →
The ring pattern is a stranger in the set, because it is not a photographic image but a computer-generated radial frequency sweep. The associated gamma value is 1, and your gamma 2.2 output is actually the correct result for gamma 1.0 processing.
And if then take a look at Your lanczos (the 2nd one, without sharpening) the same error seen in ring chart as example above, creating this "box" in center and darkening the surroundings:
The box is not an error, because a square central area is the proper response. The surroundings have a gray value of 0.5, as it should be for linear (gamma 1.0) processing.
Here is the same with ImageMagick and using gamma 1.0 (or thereabouts) to process the whole image (Lanczos resizing like in your image - Matlab Lanczos may be better than ImageMagick, I don't have any idea of Lanczos version in ImageMagick):
I assume you took the soft 16-bit original. In that case your result is virtually the same as what I get.
I find the selection of 4 of 5 photos in panels is done using only one goal in mind: finding artifacts of sharpened images.
That is not true. The buildings, yes, but the other ones are just a selection of different types of subjects. If I omit houses and buildings with roofs, vegetation, and people wearing clothes, there is not much left to choose from except barren landscapes. I was actually looking for a photo with text in it, which would just or just not be legible after the downscaling, but did not find it. I think that might be a particularly suitable subject to get a grip on original detail vs. artifacts, although I don't know whether it would work at all considering the generally small differences between the various approaches.
If I would shoot this kind of stuff it would be damn same what sharpening is used as long as it won't cause artifacts.
I don't care what the subject matter is. If there are aliasing artifacts with one subject, then the same treatment yields artifacts with any other subject even if these are not readily noticed. (Assuming that the original has frequency contents beyond the Nyquist frequency of the downscaled image.)
The only one photo had some depth of field, but it's blurry (original) and there isn't much to talk about the transition from focus to bokeh. Past years I have been evaluating sharpening and resize scripts based on "do they look the same as large non-resized or much less resized images" and "will I get same feeling of looking web photo as looking the real thing". The most important area I have been working with has been how depth of field is rendered in websize image. Typically "resize and then sharpen" will result to that the apparent DOF is multiple stops larger than when viewed in proper size.
...Show more →
That is natural, because DOF is a function of the image size at a given viewing distance.
I tried few of my scripts to your image, but none of them work since made for images, which are 5600px on long side and resizing them to 975px (or 972px recently due to Fred Miranda forum bug). This is 1:5.9 resizing ration, while in your example image we are using 1:3 ratio.
Proper handling of only a specific resampling factor is not a strong selling point.
Also the original is sharpened, and having obvious sharpening caused artifacts, which my original images won't have.
The 16-bit version has zero sharpening. If that is still too sharp, there is nothing I can do except misfocus.
You are of course free to prepare an original yourself that suits your requirements, but if it does not contain 'difficult' subject matter the issue of original detail vs. artifact will most likely be unresolved and the look-good factor is all that is left.
|
