What's the best downsizing algorithm?

Discussion in 'Digital Photography' started by Prognathous, Nov 20, 2003.

  1. Prognathous

    Prognathous Guest

    I would like to downsize a bunch of 2048*1535 photos. Which algorithm
    would provide the best results? I don't really mind the speed of
    processing as long as quality won't suffer to much.

    In addition to that, which software (for Windows) would you recommend
    that implements this algorithm and that provides batch processing? I
    would like to downsize and save the images as JPEGs, while specifying
    a size limit.

    TIA,

    Prog.
     
    Prognathous, Nov 20, 2003
    #1
    1. Advertisements

  2. Prognathous

    CSM1 Guest

    Photoshop, PSP, Photoshop Elements.

    Paint Shop Pro (PSP)
    http://www.jasc.com/products/paintshoppro/

    Try Iranview. Batch processing. Free.
    http://www.irfanview.com

    Do an experiment on one or two images. Try different algorithms, see which
    one works for you.

    There is an algorithm called "Downsize".
     
    CSM1, Nov 20, 2003
    #2
    1. Advertisements

  3. Prognathous

    eawckyegcy Guest

    All sensible approaches boil down to a resampling a suitably
    low-pass-filtered version of the original image.
    The kitchen sink of batch processing:

    http://www.imagemagick.org/www/windows.html
     
    eawckyegcy, Nov 20, 2003
    #3
  4. No matter what method/program you use you will loose detail in
    downsampling, the most obvious observation is that photos become
    "soft" so a little un-sharp mask is required to sharpen the details
    back up.

    If the originals have special effects or original artifacts such as
    grain (depending on if its a scanned blow up, or "faked" to look like
    a grainy/fast film blow up) then you will need to add this noise back
    into the photo at a greatly reduced level.

    I did have an automatic batch process set up in corel's photopaint (I
    now almost exclusivly use adobe photoshop) to re-size and "tag" all my
    web pictures, but the resulting output was just to soft and didnt
    allow visual inspection of the jpeg compression to allow for the
    smallest size files at the point just above important detail loss.
     
    Jonathan Wilson, Nov 20, 2003
    #4
  5. Prognathous

    Phil Guest

    I use Irfanview, which is free and supports batch mode. Uses a Lanczos
    algorithm and meets all my needs for straight (no cropping) downsizing
    for email and such.

    Phil
     
    Phil, Nov 21, 2003
    #5
  6. Prognathous

    Prognathous Guest

    Do you have any personal preferences? if so, which one did you choose
    and why?
    Hmm... I thought this was just a generic name for resizing an image to
    a smaller size. Which algorithm is that?

    Prog.
     
    Prognathous, Nov 21, 2003
    #6
  7. Prognathous

    Prognathous Guest

    Can you go into more details? thanks!
    I tried it, but it seems to provide no control whatsoever of the
    downsizing algorithm. Any other suggestions?

    BTW, an Italian software vendor, FSoft, claims that their similarly
    named Resize Magic provides better downsizing quality then other
    programs, including Photoshop and PaintShop Pro. Have you (or anyone
    else in this group) had any experience with it?

    Resize Magic homepage: http://www.fsoft.it/imaging/en/Default.htm

    Prog.
     
    Prognathous, Nov 21, 2003
    #7
  8. Prognathous

    Prognathous Guest

    Is there any software that can automatically apply the un-sharp mask
    after the downsizing of each image? I would like to use batch
    processing solely - for downsizing, applying the un-sharp mask and for
    saving as samller-than-100KB JPEG files.
    Thanks for the tip, but none of these photographs include any special
    effects.
    Does Photoshop provide you with better results? which algorithm do you
    use?

    Prog.
     
    Prognathous, Nov 21, 2003
    #8
  9. Prognathous

    Tore Lund Guest

    FWIW, I also wonder which is the BEST downsizing algorithm, but no one
    seems to want to answer that question.

    At any rate, the WORST routines are probably the ones included in my PS
    Elements ver. 1.01. The different resize routines of IrvanView all give
    precisely the same result when downsizing, which is indistinguishable
    from that of the bilinear routine of XnView. These routines are OK, but
    I wonder if anything better exists, because there are always some
    problems with lines that are almomst horizontal or vertical but
    differing by a slight angle.
     
    Tore Lund, Nov 21, 2003
    #9
  10. Try a small amount of Gaussian blur (depending on the implementation) 0.2 to
    0.3 per integer factor of reduction. So a 5 times reduction (20% of original
    size) would require a setting of between 1.0 and 1.5, or a ten times
    reduction would require between 2.0 and 3.0 Gaussian blur before resampling
    down. Follow this with an USM in accordance with the image content.

    The result will suffer less from aliasing artifacts.

    Bart
     
    Bart van der Wolf, Nov 21, 2003
    #10
  11. Prognathous

    eawckyegcy Guest

    Conceptually, the process is: (a) reconstruct the signal in the
    continuous realm, (b) apply a low-pass "anti-alias" filter, (c)
    re-sample at the desired rate.

    For an image, think: print the photograph onto paper, drop a sheet of
    "correctly" frosted glass over it, mount and take a new image at the
    size you want. This is all done digitally, of course.

    All the techniques can be converted into the above form. However, a
    bad reconstruction adds noise, and poor filtering produces aliases.
    Read some of your other respondents with this in mind.

    While one has huge flexibility within the above scheme, the end
    product is going to be more or less the same in the end given decent
    (a) and (b). Someone has already mentioned the Lanczos technique.
    Take a few moments at google. One hit:

    http://www.binbooks.com/books/photo/i/l/57186AF8DE

    Up and down sampling is covered in all DSP textbooks worth reading.
    I've never had a problem with the output of their resampler. But my
    resampling is usually from big -> rather small image sizes, and I just
    can't work up enough energy to be very concerned about whether there
    are some small artifacts in an N-hundred x N-hundred size thumbnailish
    images.
     
    eawckyegcy, Nov 21, 2003
    #11
  12. Isn't that a strange way? ..to damage the pictures to not damage them?

    It seems up side down.


    Morgan O.
     
    Morgan Ohlson, Nov 21, 2003
    #12
  13. I have not tried myself, but I have read that the freeware XnView is able to
    do batch-resizing and filtering.

    Morgan O.
     
    Morgan Ohlson, Nov 21, 2003
    #13
  14. QED. Try the following:
    http://www.xs4all.nl/~bvdwolf/main/downloads/Rings.gif
    is a 1000x1000 pixel 1.21MB file with concentric sinusoidal waves, with an
    increasing period towards the edges. View it at a 100% zoom setting in your
    photoeditor. Downsample this to 10% of its original dimensions
    (100x100pixels). Aliasing in its extreme is the result.

    Now do the same, but first apply a small amount of Gaussian blur. Try
    different amounts like 1.0, 2.0, 3.0 so you can see which works best with
    your editor.

    Of course this is an extreme example with large areas of almost the same
    detail, but the same happens on sharp edges, only less obvious. You would
    get the same effect on a picket fence or brick wall shot at an angle from
    such a distance that the detail approaches or exceeds the limiting
    resolution.

    Bart
     
    Bart van der Wolf, Nov 22, 2003
    #14

  15. What's your overall conclusion?

    .....I mean if people (me) don't like to take the time to try it out.

    Is it allways better to blur before compression?


    Morgan
     
    Morgan Ohlson, Nov 22, 2003
    #15
  16. Prognathous

    eawckyegcy Guest

    Apparently it is easy to miss in the N options for ImageMagick -- I
    had to carefully peruse the manpage. "-filter ...." specifies the
    reconstruction filter when resizing images. The default is Lanczos,
    and there are many others to choose from. If you have been reading
    this thread more carefully than I had read the manpage, you now know
    you can add more low-pass-filtering with the blur functions in
    ImageMagick as well. My experience with IM is that this is not
    necessary though: the defaults do a perfectly good job. Which is
    what makes IM an excellent tool: if you have to customize every last
    image, you lose most of the advantages of batch processing...
     
    eawckyegcy, Nov 22, 2003
    #16
  17. SNIP
    Experience is the best teacher...
    If the subject contains sharp edges, or repetitive structure, yes.
    How much, that depends on your photo editor implemented Gaussian Blur.

    Bart
     
    Bart van der Wolf, Nov 23, 2003
    #17
  18. Prognathous

    Mick Ruthven Guest

    BreezeBrowser (www.breezesys.com) does a great job of that. You get to
    specify the reduced pixel size, actually the maximum width and height which
    is way you want to specify it, and the sharpening specifications in a manner
    similar to Photoshop's Unsharp Mask, amount, radius, and threshold. BB also
    has a very good and very flexible website creator in which you get to
    specify the pixel size and sharpening of both the main images and the
    thumbnail images. BB's main claim to fame is as a convertor of Canon RAW
    files, so it's not well known outside of those circles, but it has many
    other very valuable and well-thought-out functions than converting Canon RAW
    files.
    ..
     
    Mick Ruthven, Nov 23, 2003
    #18
  19. If time is included, it's probably also the slowest one.
    Ok.

    Morgan
     
    Morgan Ohlson, Nov 23, 2003
    #19
  20. Prognathous

    Prognathous Guest

    According to some (e.g. JASC/PSP), Bilinear is better for downsizing
    than Bicubic. The following page provides some samples:

    http://nickyguides.digital-digest.com/bilinear-vs-bicubic.htm

    The two definitely provide very different results, but I still haven't
    made up my mind yet (each algorithm has its own drawbacks).

    Prog.
     
    Prognathous, Nov 26, 2003
    #20
    1. Advertisements

Ask a Question

Want to reply to this thread or ask your own question?

You'll need to choose a username for the site, which only take a couple of moments (here). After that, you can post your question and our members will help you out.