# Why should an image be blurred using a Gaussian Kernel before downsampling?

I recently read that before downsampling an image, it should be blurred using a Gaussian Kernel. This way, the downsampled image is better than just picking a single pixel out of a NxN block or averaging over the block. After searching in this site as well as google, I didn't get any exact answer.

But there were questions on how to select $$\sigma$$ for blurring. Reading the answers on those posts, I learned that downsampling has to be done in accordance with sampling theorem. Downsampling without blurring causes aliasing effects.

1. Can someone please explain why image has to be blurred before downsampling? I mean what is the exact relation to sampling theorem. What happens when an image is downsampled without blurring? I mean what are these aliasing effects? How can I notice them in the downsampled image?
2. Why is Gaussian blurring better than Averaging over a block?

If you can give some examples of the images, I would be a lot grateful. Ans, I would appreciate all kind of answers, partial, intuitive, rigorous, anything.

Thank You!

An image "should not be blurred using a Gaussian Kernel" in general.

This however can be a safe bet for a lot of basic image processing needs, and a smoothing is almost mandatory when you want to control the information lost by the downsampling.

Blurring is (often, not always) another word for low-pass filtering. When an image contains high-frequency content (fast variations), downsampling without blurring can produce visually weird, annoying aliasing artifacts or moiré patterns. There is an example on wikipedia Aliasing: the original

and the downsampled, represented at the same size:

The ripples on the bottom right are low-frequency artifacts generated by a careless brute-force downsampling. A blurring would attenuate image sharpness, dim the borders between bricks, and reduce the apparent aliasing aspects.

The choice of appropriate blurring filters has a long history in image processing. Gaussian shapes have long been considered somehow optimal for different reasons for "theoretical" continuous images. Plus, it is both decreasing and symmetric in the space and the frequency domains. In the time domain, this means that faraway pixels have less influence. In the frequency domain, frequencies are reduced monotonously from low to high.

Since most images are discretized, reality is somewhat different. Since the Gaussian convolution used to be computationally expensive, early approximating filters were designed with short support, borrowed for instance from Pascal triangle. Later, fast recursive implementations were designed (Deriche, Shen, etc.)

I guess question 1) is answered. For question 2) simple averaging gives an equal weight to all pixels in the window. Hence, faraway pixels are given equal importance with respect to closer pixels, which is not optimal in regions where images exhibit weak stationarity, like trends, edges and textures.

• Great answer! I’d add that simple averaging causes phase reversal for certain frequency bands, and has a much worse attenuation for the frequencies it is supposed to suppress. See here for a good example of phase reversal: crisluengo.net/index.php/archives/22 Oct 20, 2018 at 16:30
• Let me check and further read on phase reversal, I think I never thought about it that way Oct 20, 2018 at 16:38
• It comes about because, in the frequency domain, you have regions where the kernel is negative. Oct 20, 2018 at 17:28
• Gaussian kernel is separable, too, so it is computationally cheaper (You can do it on X and Y dimensions separately, and it retains its radially symmetric shape, unlike any other filter.) Mar 23, 2021 at 14:17
• I indeed did not stressed the computational aspect enough, I will be more precise. The separability is not so simple IMO : optimized integer approximations, or oriented/skewed 2D Gaussian filters are not generally separable. And some recursive implementations (like the Canny-Deriche version) were a bit spikier at the tip Mar 23, 2021 at 16:00

According to (digital) sampling theorem, signals should be properly bandlimited, before they are (down) sampled.

A practical digital filter approximately limits the bandwidth of the signal and makes it sufficiently suitable for downsampling with tolerable aliasing.

A Gausssian kernel is very suitable as a lowpass filter, as it has a number of nice features. The Gaussian function is mathematically tractable. It has sufficient frequency attenuation. It has small time domain footprint. It has little noticeable artefacts. Therefore it's kind of a programmers choice in most typical image processing applications.

• I think that Gaussians are preferred by image scientists because of their time/space duality, 2d separability and analytical tractability. For image processing that will be looked at by users, I think that other filter prototypes tends to be preferred (typically «sharper» at the cost of more aliasing and pre-ringing, and separable at the cost of rotational «nonuniformity») Mar 8 at 11:59
• @KnutInge yes so far. I do not personally prefer a Gaussian filter for any reasons but (it generates) a smoother output... Mar 9 at 17:49