The task of noise and resize

Noise and Image Size

Noise and Image Size

The task is very simple: reduce noise with downsize (downsize, i.e. downsizing).

Related issues:

  1. Is it possible to reduce a photo by 100 megapixels to 10 megapixels (10 times) and still get a 10 times cleaner, noise-free picture?
  2. Is it possible to leave only useful pixels-information in the picture when downsizing?
  3. Is there specialized software for such manipulations?
  4. How to make noise reduction during downsizing while maintaining an acceptable amount of detail using classic tools in Adobe Photoshop and other common editors?
  5. Is it possible to shoot at a high ISO, halve the image and get a noise level corresponding to ISO / 2?

Perhaps someone in the comments to share their method or reveal their secrets and tricks related to noise and downsize.

The material was prepared by Arkady Shapoval. My Youtube channeland Radozhiva's group on Facebook и VK.

Add a comment:

 

 

Comments: 53, on the topic: The task of noise and resize

  • Denis

    1. if the purity of the picture could be quantified, it would be possible to say how many times the picture became cleaner
    2. all pixels are useful, so if you reduce the size, the number of useful pixels will decrease :)
    5. No. In general, it is not connected, only indirectly

    • Michael

      You can even quantify it using Photoshop

      • Novel

        How? What is a signal and what is noise, we can only determine empirically. After digitization, everything is a signal. Take the extreme case, we photograph the matrix’s own noise to place the photo on the site - so the noise becomes a signal.

        • Michael

          Shooting worlds with uniform solid colors and brightness. Next, in Photoshop, we transfer the histogram to the brightness mode and calculate mean / std.dev - this will be a quantitative measurement of the brightness noise. No problem

          • Novel

            For a specific picture.

            • Michael

              For a specific matrix and lighting conditions. The noise level will be reproducible. It will not be exactly the same, but the numbers will be within the measurement error

              • Onotole

                I agree with Mikhail, but there is one BUT: when resizing, resizing algorithms will work out a monophonic surface and real photos in different ways.

    • Max

      And what is the problem to measure? We remove the evenly lit white sheet of paper and get a noise standard.

      • Novel

        We get a noise standard with uniform illumination of a white plain paper sheet. Real objects are unevenly lit. Noise is more noticeable in shadows than in lighting. On fine detail, texture, texture, noise is more difficult to separate from a useful signal than on uniformly colored objects.

      • Novel

        In general, noise has three components.

        Some conditionally constant associated with the heterogeneity of the sensor. Two light-sensitive cells, being illuminated by the same number of photons, conditionally falling at the same angle, will give slightly different values, because the cell size is different, a filter with a slightly different density is installed above them, etc. Those. purely physical parameter and design feature.

        Conditionally variable component, the error of each subsequent measurement.

        Well, separately, a random variable, depending on the general level of illumination. Photons are recorded as particles. Their number is finite. There is little light, the number of reflected and registered photons will vary, and the more, the weaker the level of illumination. A characteristic granularity will appear, which is fundamentally unrecoverable and the most noticeable. Just because 10 photons came from one point of the image, and 15 from another. And, given the filters, it appears in the shadows as colored noise. Or attempts to raise the exposure of a dimly lit scene. This type of noise can be eliminated either by increasing the shutter speed (the number of registered photons is greater until the noise from the matrix inhomogeneity exceeds the noise from the lack of photons), or by stacking and averaging.

        • Max

          There is a certain maximum noise level, for example, let's take 10%. Some cell showed the color components absolutely accurately, some were wrong by 5%, some by 10%. The value is absolutely random from frame to frame, even on one cell (if it is not broken). It is impossible to get something accurate from something random, except perhaps to average out of many samples, which is usually done - 10 identical frames and the average value of them - then, assuming that the noise level is still less than the maximum, noise is removed ...
          A white sheet allows you to find out this maximum level, that's all.

      • ARIES

        Gray sheet of paper ... or rather

      • Alexx

        Thus, you can only get the standard of dust on the matrix, and not noise.

  • Michael

    I'll try to answer what I was looking for)
    1. No. Noise is not proportional to resolution. This largely depends on the downsize algorithms used.
    5. In my test shots (for refueling in noise reduction), the experiment showed a decrease in ISO / 0,1 noise (roughly). So it will not work.
    For example, I took a Nikon D300 image at ISO 3200 and reduced the size by 4 times for linear resolution (i.e. 16 times) Using Perfect Resize and the standard Photoshop algorithm with bicubic interpolation. For two of the dies (light-shadow), he calculated SNR and then drove the reduced copies into Neat Image to estimate the noise level by frequency. The result is a noise level corresponding to approximately ISO 2000.
    The graph is here: https://yadi.sk/d/X5VM9FO-zTChF
    SNR Results:
    ISO 3200 2.98
    ISO 3200 downsize 6.13 (Perfect) and 5.01 (Photoshop)
    ISO 2000 6.19
    ISO 1600 7.25
    Somehow

    • Arkady Shapoval

      Very interesting

      • Michael

        Not consistent with your observations?))

        • Archie

          Arkady said: the task is simple. You have failed. Need to experiment more.

        • Arkady Shapoval

          no, really interesting observations. Is consistent

    • Eugene

      5)
      So divide by 0,1 is multiply by 10

  • Eugene o

    a) A halving of the size implies that the number of pixels on each side will decrease by a little less than a third. If you reduce the size by 50% on each side, this is equivalent to reducing the frame by four.
    Those. for a frame size of 3600x2400 (8.6 MPix) half (4.3 MPix) this will be somewhere around 2550 * 1700. In this case, the algorithm will not just collapse four pixels to one, but will conduct the recount.
    b) One of the factors for increasing ISO is a linear decrease in DD. When downsampling, the DD lost during the shooting process will be additionally eaten by the algorithm, because the noisy pixels will be diluted with normal ones (in the base scenario). Here, obviously, you need to understand the intricacies of the algorithms of a particular editor.

  • Andrew

    https://topazlabs.com/ai-bundle/

  • Oleg

    I have a question about this topic. For example, we take a full Nikon frame with a resolution of 24 megapixels and ISO 1600 and a focal length of 28mm, we take a picture. After that, we put the camera in crop mode and take an image at 18mm with the same ISO 1600. The question is how to reduce the resolution of the image on the crop and will the ISO 1600 be still clean as in the full frame?

    • Novel

      In crop mode, the camera will bite out the central part of the image that matches the same size of the crop sensor. The resolution will decrease by 1.5 times on each side, instead of 6000 × 4000 it will be 4000x2667 or whatever. The difference in noise will be insignificant, the matrix is ​​the same, the shooting conditions are the same. In terms of image quality, the difference will depend on the quality of the lens (or the quality of different focal lengths for zoom).

      • Onotole

        But visually there will be a little bit more noise, provided that both images are viewed in full screen, simply because the second image will be enlarged more than the first, and therefore all defects (not only noise but the same chromatic aberration) will also be increased.

      • anonym

        Nothing will come of it, there will be 1.5 * 1.5 times more details and noise, take a look dxo

        • Arkady Shapoval

          less details

  • anonym

    1. No. Ten times will not work. It will turn out three times.
    2. All pixels are useful. When downsizing, the developer’s algorithm works, and the more advanced the editor, the more information it will save.
    3. Topaz. Only he has a whole set of tools for each of the manipulations. For each separately. It can work from under Photoshop. Photoshop itself is not capable of high quality; it is a universal tool.
    4. No way. Classical instruments will do it not very high quality by modern standards. If etalol is not needed, then it can be done as usual.
    5. It is impossible. It doesn’t work like that. This question was already asked by experts and came to the conclusion that, until certain resolutions, one formula works to reduce noise at high ISO values, and starting with multi-pixel cameras is completely different. And in this case, the pixel density and matrix size (4/3 1.5 FF MF) plays a paramount role.

    • Arkady Shapoval

      Topaz has one problem - it is very greedy for computer resources, it is not so easy for them to drive 1000 photos, at the same time, resizers work thousands of times faster.

      • anonym

        Yes, not a hasty editor. However, over time, the algorithm will be finalized, everything goes to this. And noise is not always necessary for streaming photos.
        And how great sharpen-ai works. Just use it after de-noise and downsize. The main thing is to understand what and how to “wind up” for a good result.

    • Max

      1. This is not possible in principle.

      • NEO

        Life shows that anything is possible. The same AI can simply redraw bkb to draw a picture, draw one pixel instead of ten.

        • Max

          Well, perhaps with the help of AI ...

  • ARIES

    remove matrix noise = 50% MORE than those 12mp we make at 18mp the size with recording in 48bit tiff .. the noise will go away practically ……………… ..

    • Arkady Shapoval

      what?

      • Andrei

        I will assume that the following was meant:
        Tell a camera that can shoot at 18MP, shoot at 12MP. And the number of “noisy” pixels will go away by itself.

        In principle, this is a solution: since the camera will see more optically and process less through digital processing. Will not "suck" extra pixels (noise) out of insufficient lighting.

        And as for the digital solution, as described in the task: technically it can be realized only through AI, but I'm not sure that modern AI has reached the level to adequately determine the “signal” from “noise”.

        Another way to technically solve the problem is to do the same as it is done “inside” phone cameras: take several frames before and after and somehow compose them - the noise will be distinguishable, because on each frame may be slightly different. But again, I will not suggest software that would solve this problem in this way ...

        • anonym

          This is already being done on smartphones.
          64MP Sony matrix in the end produces a 12MP frame.

          • Iskander

            And combines adjacent pixels into one, thereby increasing the area / photosensitivity of the pixel? If yes, then this is cool, but most likely it just sprinkles the matrix or reads pixels in a checkerboard pattern.

  • Iskander

    If the photo does not require a ringing sharpness, I go to: Filters - Noise - Retouch (from one to three times) in Photoshop and only then I resize with bicubic interpolation (I did not notice much difference more clearly or smoother).

    • Arkady Shapoval

      I use third-party plug-ins for Photoshop, but none of them can reduce photos with the target in the first place to reduce noise.

      • anonym

        They are not there because downsizing does not remove noise on its own. See below for an explanation in English.
        Therefore, it is logical that the noise is removed by the programs created for this, and the downsize programs do exactly the downsize.

  • anonym

    This issue has been discussed for a long time and there is a conclusion including statistical calculations, examples and analysis.
    Simple answer
    Downsampling does not reduce noise. It's provable experimentally because it's proven mathematically.
    The noise is reduced at exactly the same rate as the detail. That's what a statistical analysis says must happen.
    It's not a way to reduce noise because you've also reduced detail. When you move away from a noisy image, you see less apparent noise (even though all the noise is obviously still there) and you also see less detail. The signal to noise ratio is still the same.
    Then you haven't reduced noise because you haven't increased the signal to noise ratio.
    In short - the signal-to-noise ratio does not change when downsized, noise is removed along with the details.
    We need to reduce the noise - we need to use software and not downsize.

    • anonym

      Rave. Downsize visually reduces noise, I’ve been using it for 10 years. Mathematically, the signal noise may be the same, but look 1 to 1 full size from a noisy picture and a reduced size of 1 to 1 and everything will become clear

      • anonym

        Before stating that something is nonsense, take the trouble to carefully read what is indicated.
        Hint - if you reduce the frame from 24MP to 0.5MP, for example, the noise will obviously be less as well as useful information, but the signal / noise will remain unchanged.
        And if reduced to a point, there will be no noise as well as useful information.)
        Because downsize removes all information and does not reduce noise.

      • anonym

        Your example has long been discussed a million times.
        No information - no pixel - no noise.
        And for a long time there is another example where the noise reduction program is used and compared with the frame after downsizing. And the difference is visible when the frame without downsize but after special noise reduction is visually much cleaner and with details in comparison with the frame after downsize.
        Because cutting / reducing frame downsizing is not noise reduction in principle.

  • Peter Sh.

    I suspect that it all depends on the type of noise. Well, the noise is noisy. Also, different editors crush noise in different ways.
    Of all the editors, Rawtherapy makes the best noise. But there you need to be able to dance with a tambourine.

    There is a certain limit after which noise becomes unbearable. And then nothing will come of it.
    In short, I do not think resize is an effective tool against noise.
    In general, the noise is not so terrible as it is painted. The viewer does not see noise at all.

  • Alex

    Interestingly, is it possible to use the principle of on-camera noise reduction from Kknon at long exposures in the PS for “cleaning” the noise?
    A picture is taken with the shutter open, followed by a picture with the shutter closed with the same pull, and then subtracts the noise from the first in the second.

    • Michael

      In the forehead does not work. You can stack multiple takes

  • Alexey

    From personal experience. Perhaps this is relevant to the topic.
    The "inch" Sony Cyber-shot DSC-RX100M4, which already makes noise at ISO 400, has an anti-shake scene mode (Anti Motion Blur), which allows you to shoot at high shutter speeds (1/100 and shorter), but ups the ISO to 6400 And here's the thing with him. In the menu I reduce the photo size from 20 Megapixels to 5, take a photo, there is a second delay for processing, and the output is a clean and sharp picture. How it's done - I can't imagine. This mode has become one of my favorites.

    • Iskander

      But really, it's interesting to me too! Many people recommend shooting at maximum resolution, although it is not required, and then doing magic in editors. Hence the question for professionals - if the menu is set to medium or low resolution, how does the camera reduce it? Combining neighboring pixels into one (with a larger area) or "pulling out" information in a checkerboard pattern? Or, more specifically, does the pixel area increase when going from quality L to quality S?

  • Eugene

    From personal experience, I print 10 × 15, very rarely A4, I look (show) on the tablet 10′-noise, I do not see it, it’s too extreme. (Canon 60d, iso <3200) 😆

    • Iskander

      That's right, it's the same downsize, but in print. In many cases, noise can only be seen when viewed pixel by pixel on a monitor or when printing posters. If it is visible even at 10x15 - then byada (. Here, only the translation to B / W and the addition of the filter "film grain" will be saved, ala retro, like it was conceived).

  • Andrei

    “Can a 100 megapixel photo be reduced to 10 megapixels (10 times) and still get 10 times cleaner, noise-free picture?”
    No, you can’t, I will write about this at the end of the message. But, as you know, there are noises - the consequences of debayerization. There will be less such noise compared to shooting with a 10-megapixel matrix.

    "Is it possible to leave only useful pixels-information in the picture when downsized?"
    You can transfer a little more details to the reduced version if you add sharpness before reducing (I add sharpness only in the brightness channel so as not to amplify color noise). But not reducing all the details while decreasing. Adding sharpness after reduction is already just a special effect and will not bring new details.

    "Is there specialized software for such manipulations?"
    I use GIMP, the LCh Color, LCh Lightness channel modes, Unsharp Mask filters, Selective Blur turned out to be useful, these are my main “tools” for noise reduction. Alas, there are no such channel modes in Photoshop, but there is Lab there and you can achieve similar results.
    I’m sure to switch to 32-bit color (or at least 16-bit) in order to get a higher color bit and fewer artifacts after subsequent reduction after reduction. By the way, here's what really can be obtained from a decrease so an increase in the color bitness (depth).

    "How to properly reduce noise while downsizing while maintaining an acceptable amount of detail using classic tools in Adobe Photoshop and other common editors?"
    I didn't notice the difference between “suppress noise, then reduce size” and “decrease size, then suppress noise”. Unless on the enlarged version, the noise reduction radius should be larger and it takes a little more CPU time. But I suppress only color noise using an algorithm like Selective Blur (I make the threshold a little smaller, the radius - depending on the noise level and image size), occasionally I can overlay a 50% layer with suppressed noise in the brightness channel if the brightness noise is too striking. Sometimes you can select non-critical areas (blurred backgrounds) with a magic wand and work with them separately.
    But in general, I do not consider it necessary to remove luminance noise, and in principle it is impossible to remove it so that it remains invisible. The same Topaz does not “remove” the noise, it draws the details absorbed by the noise and sometimes does it very roughly and unnaturally, sometimes thinning the thicker lines, leaving patterns characteristic of neural networks. It's better to put up with luminance noise as well as with grain on the film.

    "Can you shoot at a high ISO, cut it in half and get a noise level that's ISO / 2?"
    Based on my experience: no, noises will be noticeable on a reduced copy, and here's why:
    the nature of matrix noise is such that noisy pixels do not neutralize each other at all, as you suggested by your questions, but simply merge into larger clumps of noise. So you have to deal with it and do it quite simply (but the colorless grain will still remain and there is nothing wrong with that).

  • kotofei

    And the problem is of practical importance for Arcadia ... so .. July 2019, and in September sony a7r-4 came out. I will rephrase the problem: “is it worth switching to A7-4 and doing subsequent resizes all over the place, incl. in order to reduce noise "=)

    I propose to look at this from a different point of view. Are you shooting a frame to get which product in the end? A4 fingerprint or viewing on a 4k monitor / TV? Well, in most cases, probably so. We rarely print. 8k TVs will not be mass soon, so the assumption of 4k TV is quite appropriate, right? 4k TV is 3840x2160, but it is 16: 9 and we shoot in 3: 2, i.e. AT MY LOOK, after developing and retouching, you need to resize the product to 3240x2160, and this is just a little less than 7 megapixels.
    By the way, this indirectly confirms the uselessness of multi-megapixel sensors.

    PS topaz gigapixel knows how to use the resources of the video card - it counts many times faster than the processor. But still 1000 frames is a long time. And what's the difference, the main thing is that you would have time to calculate everything before the next photo session, right? =)

Add a comment

Copyright © Radojuva.com. Blog author - Photographer in Kiev Arkady Shapoval. 2009-2020

English-version of this article https://radojuva.com/en/2019/07/how-to-make-plumbus/