3
$\begingroup$

I am researching downscaling methods for image downscaling. I've both read and found area-averaging to work well in Python using CV2, but the settings in MATLAB allow for it's bicubic interpolation downscaling to work better.

Are there any resources comparing image interpolation techniques? especially explaining the theory? (I can't seem to find any books or journal articles that cover what I need).

Thanks

Improve this question
$\endgroup$
6
  • 2
    $\begingroup$ This is screwy enough that I'm only putting it out as a suggestion, not an answer. Gimp (and image majick, upon which it is based) is open-source, and does an excellent job of downscaling images (and upscaling, for that matter). At worst, you can download the source from their git account and look. $\endgroup$
    – TimWescott
    Commented Apr 12, 2021 at 16:26
  • 1
    $\begingroup$ A thumb up for GIMP and ImageMagick, with a lot of command line and batch possibilities $\endgroup$
    – Laurent Duval
    Commented Apr 12, 2021 at 17:02
  • 1
    $\begingroup$ @TimWescott same for OpenCV (i.e. CV2), it's open source and has readable source code. Also, I think every textbook on image processing would cover scaling in an early chapter, as it's such a basic operation in many alorithms. $\endgroup$
    – Marcus Müller
    Commented Apr 12, 2021 at 17:33
  • 1
    $\begingroup$ which type of images: computer generated (non bandlimited) or natural (bandlimited)? $\endgroup$
    – Fat32
    Commented Apr 12, 2021 at 17:51
  • $\begingroup$ @Fat32 Natural images. $\endgroup$
    – Dalton
    Commented Apr 12, 2021 at 20:22

2 Answers 2

Reset to default
1
$\begingroup$

Image scaling is a variation on the dsp topic of resampling. A text book on resampling to a lower rate tells us that a lowpass filter that remove all signal energy beyond 1/2 the sampling rate of the lower rate will completely avoid aliasing. Making such a filter without disturbing the passband means an ideal lowpass filter, something that cannot be practically realized.

Thus, a traditional image scaler (or resampler) can be interpreted as various trade-offs in linear lowpass filter design so as to minimize aliasing, maintain a flat passband and avoid pre-ringing.

Anything concerning spatial phenomena targeted for our vision tends to use really small kernels. Ie «bad» frequency-domain selectivity. Often, people prefer some residual aliasing rathet than loss of sharpness. So a compact lowpass filter prototype with only moderate attenuation at 1/2 the sampling rate might be a decent place to start

Improve this answer
$\endgroup$
0
$\begingroup$

There is a great analysis on Juan Conejero, PTeam - Interpolation Algorithms in PixInsight. Usually you will find such comparisons in photography communities. Probably there are better methods which take edges into account.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.