Detailed upsampling instructions, please

Discussion in 'Digital Darkroom' started by chris_newkumet, Nov 3, 2004.

  1. Forgive the basic nature of my question, but I really have searched
    high and low for step-by-step instructions for upsampling using the
    10% increment approach in PS/bicubic smoothing. I even dug through
    Real World Photoshop CS, to no avail. I have a .tif image currently
    sized at: 3072x2048 pixels; 17.067"x11.378"; 180 resolution. The
    images is exposed well and is reasonably sharp. I need to make a
    cropped 10"x20" print with resolution of 250 dpi, with sharpening
    done as the last step. In cropping, I won't be taking much from the
    sides. Assuming I only want to use PS (which I do, so no Genuine
    Fractals advice, please), the common wisdom seems to be to do this
    through the document size window, stepping up 10% at a time with
    bicubic smoother turned on. Ok, exactly how do I do that to reach my
    desired print size? Again, sorry to be dense, but I mean exactly.
    Chris N.
  2. Forgive me for not directly addressing your question. But I have formed the opinion that folks here sometimes engage in what could be called superstitious behavior. Someone, somewhere, decided that upresing in small steps was better than in one step. But I doubt that that person, or the others who have followed that advice, have made a mathematical determination of which was genuinely better. Although I'm not an expert in image processing, I've taken a few graduate classes, and for the life of me I can't figure out how this particular meme caught on (and why 10% at a time, in particular, anyway?). It's just goofy. My recommendation is to use bicubic smoother in one step, see how that looks to you, and if you are happy, be done with it. Cheers.
  3. Hi Chris, Fred Miranda might have been one of the first to recommend "stair interpolation". See:

    I can't say for sure if there is any benefit to increasing image size incrementally; however if you use Photoshop's Actions feature, it's not much more trouble than doing it in one fell swoop.

    Here's how I do it:

    1) Select Image>Image Size

    2) Check the box that says "Resample Image"

    3) Using the dropdown boxes for Width and Height, change them to use percent instead of pixels

    4) Now change the %100 percent figure to, say %110 to increase the image size by 10 percent and press Enter or click OK.

    That's it. As I mentioned above, you might want to use Photoshop's Actions feature to record this. You can then "play it back" with one keystroke. For example, I've assigned my function key F11 to increase my image size by 10 percent using this method.

    Good luck!
  4. I once did an admittedly informal test. I took an image that I wanted to up 3x. First I upped it in one lump using Photoshop. Then I upped it in 10% increments. Then I upped it in one lump using GenuineFractals. The one lump jump in photoshop did in fact look worse than the other two options--there was a real loss of detail noticeable. The one jump up in Genuine Fractals I found to be as good as the 10% steps in photoshop. So, my conclusion was that the 10% stair method does work better than the one jump up in photoshop, but GenuineFractals does just as well as stair step method in one jump.
  5. Was your testing with Photoshop 7 standard "bicubic" or the newer "bicubic smoother" in Photoshop CS?
  6. B.Hooker's directions are correct, but they didn't *quite* answer the question "... exactly how do I do that to reach my desired print size? Again, sorry to be dense, but I mean exactly."

    First off, you're starting with a 3:2 ratio image (3072:2048) and want a 2:1 when you're done --- 5000 x 2500 per your description. I'd crop the image first, constraining the image to a 2:1 ratio using the toolbar. Not 2in/1in, not 2px/1px, not 5000px/2500px, just 2/1. That will chop your image down the the desired aspect ratio without resampling.

    After that, follow B.Hooker's description until the image just barely goes over 5000 x 2500. Do one undo (ctl-Z on a PC) and it'll within 10% of your final size, but just under it. Now resample one last time, but don't do 10% relative, but enter resolution 250dpi (first) followed by 5000pixels by 2500pixels under "Pixel Dimensions".

    I do agree with other posters that using the new "bicubic smoother" in a single step is now the best method. I also think there's about a 99% chance that your printer uses a better algorithm for printing your image at 10"x20" so you'd be even better off just cropping it, leave the dpi & dimensions alone, and handing it off. Some people *do* enjoy the busywork though! :)
  7. Upsampling (interpolation) in Photoshop CS is plain and simple. Use Bicubic Smoother, one step. Period.

    Forget about the FM thing, 10% increment, those were pre-Photohop CS.
  8. I doubt there's anything magic about 110%.
    You could use 109% or 111% and get similar results.
    The tradeoff between Stairstep and Bicubic Smoother
    is preserving detail versus producing straight diagonals.
    Image content dictates.
    I would recommend trying one-step Bicubic Smoother,
    then four 113% repeated, and comparing results.
    Because 3072 * 1.13 ^ 4 = 5009, you will then be able to crop
    9 pixels from the sides.
  9. Stair interpolation (SI) simply iterates an interpolation function to effectively simulate using a higher order polynomial approximation of the underlying image. It is not a well thought out practice and it tends to create artifacts in your image (i.e., it makes up details that were not there). Stair interpolation mixed with a bicubic interpolation method tends to create a result very similar oversharpening an image or heavily truncating a Fourier Series or using moderately heavy JPEG compression. And the primary artifact it generates is ringing (haloes) about edges. As an aside, heavy JPEG compression does not have this problem because blocking artifacts tend to obscure the ringing.
    Below is an image (blown up 200% to show more detail) illustrating the effects of various interpolation techniques in PS to increase the resolution of an image by 194.5%. This is the equivalent of using seven 10% steps in SI.
    A good starting point to better understand this for those who finished linear algebra or advanced calculus in college can be found at wikipedia. For those who chose not to study such topics, just take a look at the graphs on this page and note how the shape of a polynomial of higher degree (what SI essentially creates) can get wavier and wavier. Specifically, note how a polynomial of degree 3 (i.e., a cubic polynomial) is less wavy than a polynomial of degree 5. While these examples are 1 dimensional, the underlying basics of polynomials of the same degree in more dimensions (2 dimensions for a photo) are qualitatively similar. For those interested in further understanding, you might look into splines and numerical analysis. Some additional understanding of Numerical stability will begin to help you understand why round off errors in 8-bit color are ameliorated by working in 16-bit color.
    Hence, beyond simply doing your interpolation in a single step, I recommend upressing an image in 16-bit color mode as you will get smoother gradients (and less combing of the histogram). If you want sharpening halos or noise added to your image, then doing such yourself will give you more control over how your image looks.
    As for Fred Miranda, I have only seen his "actions" attributed to poor quality work and I would suggest avoiding them. People often say, "it cannot be oversharpened/badly done/... because I used an FM action to do it." Sadly, simple or complex tools will never be a good substitute for using your mind if you care about the quality of your work. And one size fits all answers rarely fit anything well.
  10. And if higher-order poynomials are what you want then you should proabably just find yourself a program that implements lanczos and other more advanced interpoaltion algorithms.
  11. Sean, it's interesting that Nearest Neighbor produces the best
    results in your test example, but it usually produces the worst
    results with real-world images. Have you seen Bart van der Wolf's
    concentric rings test?
  12. The problem with the example is that the edges between the blocks are square. Square waves are _the_ pathological case for spline interpolation. (I.e. they produce the highest possible degree of distortion.) If the starting point were low-pass filtered first (even a little gaussian blur would help but something stronger and less smeary would be better), it would be more representative of real-world images.

    I agree that manually stepping is preferable to an action. That way the step width can be adjusted to the image (and in particular to its starting contrast) -- and halos can be effectively suppressed using input/output levels. I haven't experimented much with the new bicubic functions, but with the regular one I would step once 50-100% to get within ballpark, then take a few shorter more careful steps to get to the final result.

    These days though I tend to work more from scanned film, and all interpolation I do is aimed at going down to print. The few digital shots I take are for web use (with my 3MP D30). So haven't had much reason yet to figure out how the new bicubic functions can help. My experience with the FM actions though is that they work quite well for real uses. But SI tends to be a little too clinically clean for my personal taste, I like Genuine Fractals personally once the images go beyond 200%. (At least Canon DSLR ones. But they're not crispest on the planet to begin with.)
  13. Man, my grammar just stinks today. Gotta do something about this keyboard, it just can't type for sh*t. And I'm not sufficiently caffeinated to properly supervise it yet.
  14. Sean, it's interesting that Nearest Neighbor produces the best results in your test example, but it usually produces the worst results with real-world images. Have you seen Bart van der Wolf's concentric rings test?
    Bill, the aboive is a a simple example that was contrived to show the one dimensional behavior of these methods. By illustrating the one dimensional behavior the examples then tie into the polynomials that are taught in high school mathematics courses. This in turn yields a situation (shapes with sharp edges and square corners) that gives us an ideal case for the NN (Nearest Neighbor) resampling method. But it also clearly illustrates how the other sampling methods introduce new data into the image (i.e. interpolate intermediate values) and how those new values do not necessarily conform to the original data.
    Included below is another simple example that include 45 degree and 75 degree lines so that aliasing can be seen. It is clear from aliasing the NN is not ideal and the the blurring of the other methods can have beneficial effects in interpolation. I also include an example where you can see that SI can create truly pathological nonsense when mixed with NN resampling.
    Beyond that, I vaguely rembering seeing Bart's examples at some point in the past. The concentric rings here create another pathological case. You should try stair interpolation on Bart's test image if you want to see some really ugly behavior. Whereas, real world images tend to land somewhere in between these two extremes and tend to add color to the mix.
    Which is better, which is worse? In the end it depends on the image. Personally, I just use Bicubic Smoother and some sharpening to suit the output device. Feel free to make your own choice. Experiment with the type of images you shoot. I personally do not like the extra ringing effect about sharp edges SI introduces over the equivalent single step bicubic routine. Nor do I care for the way SI tends to smooth low frequency textures (this reduces apparent noise, but it also smooths away detail). But that is just me.

Share This Page