I`m reading about Bilateral filter with images, I knew that it uses Gaussian filter (which is a LPF) as a domain filter plus a range filter and it is used to preserve edges(high freq component).

Can we consider Bilateral filter as a LPF ? .

I searched some articles: 1- Which approach is better for decomposing a image into high frequency and low frequency component?

The answer was that bilateral filter attenuates ""medium frequency "

2- How to Extract High Frequency and Low Frequency Component Using Bilateral Filter?

The answer was to consider bilateral filter as a LPF.

  • $\begingroup$ I posted and an answer. I think the main thing here is to think in the Patch Level. $\endgroup$
    – Royi
    Commented Oct 5, 2019 at 0:25

3 Answers 3


Neither. To me, filter classes using the notion of frequency bands (low-pass, high-pass, etc.) can be used safely in the linear case. And the bilateral filter is nonlinear. Edges are not really high-frequency: they often have sharp variations across the edge, but slow variation along it. Hence linear directional filters often have a derivative part and and orthogonal smoothing part.

I would consider the bilateral filter as an edge-preserving smoother, a broad and somewhat imprecise class.

Suggested reading: Fast and Provably Accurate Bilateral Filtering, 2016

The bilateral filter is a non-linear filter that uses a range filter along with a spatial filter to perform edge-preserving smoothing of images. A direct computation of the bilateral filter requires O(S) operations per pixel, where S is the size of the support of the spatial filter. In this paper, we present a fast and provably accurate algorithm for approximating the bilateral filter when the range kernel is Gaussian. In particular, for box and Gaussian spatial filters, the proposed algorithm can cut down the complexity to O(1) per pixel for any arbitrary S. The algorithm has a simple implementation involving N+1 spatial filterings, where N is the approximation order. We give a detailed analysis of the filtering accuracy that can be achieved by the proposed approximation in relation to the target bilateral filter. This allows us to estimate the order N required to obtain a given accuracy. We also present comprehensive numerical results to demonstrate that the proposed algorithm is competitive with the state-of-the-art methods in terms of speed and accuracy.

  • $\begingroup$ What about the Patch Level? Have a look on my analysis. $\endgroup$
    – Royi
    Commented Oct 17, 2019 at 11:27

Bilateral Filter is indeed an Edge Preserving Filter.
Moreover, due to being Spatially Variant Non Linear Filter it can not be applied using Fourier Transform.
Since it has no representation in the Frequency Domain, it is not well defined how to classify it into one of the categories: LPF, HPF, BPF or BSF.

Nonetheless, let's try doing some analysis based on analyzing the filter itself and some empirical analysis.
If we look at the filter per patch defined by the radius of the filter, we have fixed weights and we can analyze its effect.

Analysis of the Bilateral Filter Formula

The Bilateral Filter is given by:

$$ O \left( i, j \right) = \frac{1}{ {W}_{i, j} } \sum_{m = -r}^{r} \sum_{n = -r}^{r} w \left( i - m, j - n \right) I \left( i - m, j - n \right) $$


  • $ O \left( i, j \right) $ - Output image value at pixel $ \left( i, j \right) $.
  • $ I \left( i - m, j - n \right) $ - Input image value at pixel $ \left( i - m, j - n \right) $.
  • $ r $ - Radius parameter of the filter.
  • $ w \left( i - m, j - n \right) $ - Weight of the pixel $ \left( i - m, j - n \right) $ given by $ w \left( i - m, j - n \right) = {w}_{s} \left( i - m, j - n \right) {w}_{r} \left( i - m, j - n \right) = \exp \left( - \frac{ { \left( i - m \right) }^{2} + { \left( j - m \right) }^{2} }{ 2 {\sigma}_{s}^{2} } \right) \exp \left( - \frac{ { \left( I \left( i, j \right) - I \left( m, n \right) \right) }^{2} }{ 2 {\sigma}_{r}^{2} } \right) $.
  • $ {W}_{i, j} $ - Normalization factor of the pixel $ \left( i, j \right) $ given by $ \sum_{m = -r}^{r} \sum_{n = -r}^{r} w \left( i - m, j - n \right) $.

So we have the Spatial Weight, which is just classic Gaussian Filter $ {w}_{s} \left( i - m, j - n \right) $ and we have the Range Filter $ {w}_{r} \left( i - m, j - n \right) $.

Assuming we fixed the Spatial Filter with certain Radius and parameter $ {\sigma}_{s} $, let's analyze the effect of the Range Filter.

If we have $ {\sigma}_{r} \to \infty $ then the Range Filter has same value for any pixel and we basically have Spatially Gaussian Filter which is LPF.
For $ {\sigma}_{r} \to 0 $ we'll have zero weight for any pixel which is not $ \left( i, j \right) $, which means Delta Filter (Identity Filter). Namely no effect at all.

So the Bilateral Filter is behaving, per patch, as something between Identity Filter to LPF Filter.

Empirical Analysis of the Bilateral Filter on Patches

Let's take the Lenna Image and analyze, empirically, the Bilateral Filter over few patches.

The Lenna and selected patches for analysis Image is given by:

enter image description here

Let's see how the weights and the Frequency Domain looks:

enter image description here

enter image description here

enter image description here

enter image description here

enter image description here

As can be seen in the results above, the Bilateral Filter is indeed data dependent.
We chose different patches (2 Steps, Texture and Flat) and can see how it behaves for each.

If we look on the Frequency Domain, unless the rangeStd parameter is very low or the patch have high variance it looks like LPF behavior.


The Bilateral Filter isn't a classic Linear Spatialy Invariant filter.
Hence it can not be classified like classic filters.
Yet according to the analysis above one could come to this conclusion:

  1. When the Bilateral Filter indeed smooths (The rangeStd parameter is modes relative to the data variance) it behaves, at the patch level, like LPF filter.
  2. When the rangeStd is very low compared to the variance of the data in the patch the Bilateral Filter behaves almost as the Delta Filter (Identity).

The main idea here is since this is Spatially Variant filter we have to analyze it on the Patch Level and not the Image Level.

The full code is available on my StackExchange Signal Processing Q60916 GitHub Repository (Look at the SignalProcessing\Q60916 folder).


I’d say it is a variation of lowpass filters. It tries to average some neighbourhood eg in order to reduce variation caused by noise, but it restricts the contribution different from a LTI filter.

Ie having only positive weights makes it a «kind of lowpass» to my mind.

I am aware that this definition would fit with a positive comb filter, but then that depends on what range of frequencies is the main motivation for employing a comb filter.


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.