Low-pass filter parameters for image downsampling

Question

I need to downscale an image in a factor of $s_x$ horizontally and $s_y$ vertically ($s_x$, $s_y$ < $1$).

I want to use a finite $n\times m$ low-pass filter before downsampling.

How should I determine the low-pass filter parameters ($n, m$ and the Gaussian $\sigma$) to get it as a function of $s_x$ and $s_y$?

Specifically, I'm interesting in the case where $s_x=s_y=1/\sqrt{2}$.

Answer 1

You have to think about the change in Nyquist frequency between both images. If the Nyquist frequency of the original image is N, the downsampled image will have a lower Nyquist frequency, xN, where x is related to the ratio of sizes between the final image and the initial one. You would need to remove those spatial frequencies which are higher than xN in the original image before downsampling it.

The power spectrum of a Gaussian in the image space, is also a Gaussian in the frequency space. If we ignore for a moment the second dimension, the Gaussian in the image space is defined as exp(-x^2/s^2), where x represents your pixels. This is mapped to the frequency space as exp(-w^2*s^2), where w is the frequency. The sigma parameter (s) shows that a broad Gaussian in the image space, corresponds to a narrow Gaussian in the frequency space.

You would like to choose a sigma parameter that yields a very low value in frequency space at the frequency that corresponds to the Nyquist frequency of the down sampled image.

Answer 2

It's already been pointed out that $n$ and $m$ are to be chosen based on $\sigma$.

I've spent some time thinking on how to pick $\sigma$ best. Here are my considerations. tl;dr: Maybe I made some mistake, but $\sigma^2\approx3.37$ looks like a good choice for down-sampling by factor 2.

Answer 3

If you were doing a large downsize (say 2x, 3x, 4x) you can do pixel averaging to achieve good anti-aliasing. That's actually why anti-aliasing uses a lot of additional CPU/GPU in order to make video games look crisper.

Since you're going from a 1000x1000 to a 707x707 image (just an example for the scale factor) you're correct that aliasing might be an issue.

Thankfully this is a problem that many people have run into already and done quite a bit of work to solve. In many cases a bicubic interpolation is the way to go. There are some examples of what the different interpolation methods look like here:

http://www.compuphase.com/graphic/scale.htm

OpenCV's resize has several of those methods built-in:

http://opencv.willowgarage.com/documentation/cpp/geometric_image_transformations.html#cv-resize

If you have already played around with several of those interpolation methods and they don't work well, please post some kind of a sample source image and sample result image showing the shortfall. We'll need this in order to diagnose the problem and try and come up with a good solution to it.

Answer 4

I don't really have a good answer for you, but here are 2 options that you could try:

• in Computer Vision , this rescaling factor is usually handled by applying a Gaussian filter of width (in pixels) between 5 and 9. You can find the corresponding $\sigma$ of the Gaussian because the pixel width is classically equal to $3 \sigma$.
• if you are willing to do fine signal-sampling computations, then why not using Fourier transform ? Take the FFT of your image, keep only the subpart that corresponds to your target size, and invert the transform. This will apply an anti-aliasing filter on your spectrum. If there are too many artifacts (ripples, ringing) for you, then apply on your spectrum of Gaussian filter whose width corresponds to your target size .

Answer 5

For the case $s=1/\sqrt{2}$, the classic Gaussian kernel in 2D is of the form: $$[1, 2, 1]^T[1, 2, 1]$$