Smoothing & Blurring Filters

Smoothing (blurring) reduces noise and detail in an image by averaging or modifying pixel values based on their neighbors. It works as a low-pass filter, allowing low-frequency components (smooth regions) to pass while suppressing high-frequency components (noise, edges, fine details). They are almost always the first preprocessing step before edge detection or segmentation.

Classification of Smoothing Filters

graph TD
    A[Smoothing Filters] --> B[Linear Filters]
    A --> C[Non-Linear Filters]
    
    B --> B1[Mean / Average Filter]
    B --> B2[Gaussian Filter]
    
    C --> C1[Median Filter]
    C --> C2[Bilateral Filter]
    C --> C3[Min/Max Filter]
    
    style A fill:#4A90D9,color:#fff
    style B fill:#7BB661,color:#fff
    style C fill:#E8A04C,color:#fff

1. Mean (Average) Filter

Linear Spatial
Replaces each pixel with the simple arithmetic average of all pixels in a rectangular neighborhood (kernel). Every neighbor contributes equally.

K = \frac{1}{9} \begin{array}{ccc} 1 & 1 & 1 \\ 1 & 1 & 1 \\ 1 & 1 & 1 \end{array}

When to use ?

Very simple to understand easy to implement method for smoothing images.
Quick, low-cost noise reduction before further processing.
When image quality is not critical and speed matters.
Preprocessing step in simple pipelines.

When to avoid?

When edges and fine details must be preserved — this filter blurs everything uniformly, including edges.
On images with salt-and-pepper (impulse) noise — averaging will spread the outlier values rather than remove them.

Caution

The Mean Filter is the bluntest tool in the kit. It treats noise and edge information identically. Use it only when you need fast, rough smoothing and do not care about edge sharpness.

2. Gaussian Filter

Linear Spatial
The effect of Gaussian smoothing is to blur an image, in a similar fashion to the mean filter. The degree of smoothing is determined by the standard deviation of the Gaussian.

The Gaussian outputs a weighted average of each pixel's neighborhood, with the average weighted more towards the value of the central pixels. This is in contrast to the mean filter's uniformly weighted average. Because of this, a Gaussian provides gentler smoothing and preserves edges better than a similarly sized mean filter.

How it works?

Applies a weighted average using a Gaussian (bell-curve) function. Pixels closer to the center receive a higher weight than those at the edges of the kernel. The standard deviation $σ$ controls how much blurring is applied — a larger $σ$ produces a stronger blur.

A common example of 3×3 discrete approximation ( $σ \approx 1$ ):

K = \frac{1}{16} \begin{array}{ccc} 1 & 2 & 1 \\ 2 & 4 & 2 \\ 1 & 2 & 1 \end{array}

The 2D Gaussian Formula

G (x, y) = \frac{1}{2 π σ^{2}} e^{- \frac{x^{2} + y^{2}}{2 σ^{2}}}

where

$x, y$ : The spatial distances from the center of the filter mask in the horizontal and vertical directions, respectively.
$σ$ : The spread of the filter. (standard deviation)
$σ^{2}$ : The variance.
$e$ : Euler's number (approx. 2.718), representing exponential decay

Step-by-Step Example

1. Create the Kernel Weights

Let’s calculate the weights for a $(3 \times 3)$ kernel using a standard deviation of $(σ = 1)$ . The center pixel is at $(0, 0)$ .

Center pixel (0, 0):
$G (0, 0) = \frac{1}{2 π (1)^{2}} e^{- \frac{0^{2} + 0^{2}}{2 (1)^{2}}} = \frac{1}{2 π} e^{0} \approx 0.15915$
Immediate horizontal and vertical neighbors (e.g., $(- 1, 0)$ , $(0, - 1)$ , $(1, 0)$ , $(0, 1)$ ):
$G (1, 0) = \frac{1}{2 π} e^{- \frac{1^{2} + 0^{2}}{2}} = \frac{1}{2 π} e^{- 0.5} \approx 0.09653$
Corner neighbors (e.g., $(1, 1)$ , $(1, - 1)$ , $(- 1, 1)$ , $(- 1, - 1)$ ):
$G (1, 1) = \frac{1}{2 π} e^{- \frac{1^{2} + 1^{2}}{2}} = \frac{1}{2 π} e^{- 1} \approx 0.05855$

2. Normalize the Weights

If we add up all the unrounded weights calculated above (1 center + 4 edges + 4 corners), the sum is roughly ( $0.667$ ). To ensure the final image brightness remains unchanged, we divide each value by the sum of all weights. Dividing and rounding gives us a clean, integer-based $(3 \times 3)$ Gaussian kernel:

\frac{1}{16} \begin{array}{ccc} 1 & 2 & 1 \\ 2 & 4 & 2 \\ 1 & 2 & 1 \end{array}

3. Apply the Filter (Convolution)

Imagine a section of your image has the following pixel values:

[\begin{matrix} 10 & 20 & 10 \\ 20 & 50 & 20 \\ 10 & 20 & 10 \end{matrix}]

To find the new smoothed value for the center pixel, we multiply the image pixels by the Gaussian kernel weights and sum them up:
$New Center = \frac{(10 \times 1) + (20 \times 2) + (10 \times 1) + (20 \times 2) + (50 \times 4) + (20 \times 2) + (10 \times 1) + (20 \times 2) + (10 \times 1)}{16} = 25$

4. Result

The center pixel value is smoothed from $(50)$ down to $(25)$ , blending it into the background.

When to use ?

Standard preprocessing before edge detection (e.g., before applying Sobel or Canny).
Reducing Gaussian noise (random sensor noise from cameras).
Any time you need smooth, natural-looking blurring.

How to pick $σ$ ?

Small $σ$ (e.g., 0.5–1.0): light smoothing, retains more detail.
Large $σ$ (e.g., 3–5+): heavy smoothing, more detail is lost.
Rule of thumb: kernel size should be approximately $6 σ \times 6 σ$ (odd-numbered).

When to avoid ?

On images with salt-and-pepper noise — Gaussian blur spreads isolated pixel outliers into the surrounding region.
When edges must be completely preserved.

Caution

Gaussian blur is the default go-to, but choosing $σ$ too large will destroy fine details that your downstream step (e.g., edge detection) depends on. Always verify blur intensity visually or empirically.

3. Median Filter

Non-Linear Spatial
The median filter is normally used to reduce noise in an image, like the mean filter, but it does better job in preserving details in the image in comparison.

How it works?

Median filter, sorts all pixel values in the neighborhood and replaces the center pixel with the median (middle) value.
By calculating the median value of a neighborhood, the median filter has two main advantages over the mean filter:

Robust to outliers: The median is a more robust average than the mean and so a single/few outlier pixel(s) in a neighborhood will not affect the median value significantly.
Preserving sharp edges: Since the median filter's pick value must actually be the value of one of the pixels in the neighborhood, the median filter does not introduce new intensity pixel (calculated value) when the filter straddles an edge, thus the median filter is much better at preserving sharp edges better.

When to use?

Images corrupted by salt-and-pepper noise (random black and white pixels) — this is where the Median Filter excels.
When you need noise removal but must preserve sharp edges.
Applications: Scanned documents, binary images, or medical images with impulse noise.

When to avoid

Gaussian noise is characterized by small, random variations distributed across the entire image.
A median filter works by picking a single pixel value from a sorted neighborhood and if more than few pixels has a tiny amount of random noise, picking the median pixel will just selects a different noisy value.
A Gaussian filter is much better here because averaging mathematically cancels out zero-mean random noise, smoothing the image out beautifully.
When computational speed is critical — sorting is more expensive than convolution which is computationally expensive, especially as the kernel size or image resolution grows.
If you run a median filter over an image multiple times, it continuously forces neighboring pixels to take on identical values. This destroys fine, low-contrast textures (like hair, fabric, or grass) and turns them into flat, artificial-looking blocks of color.

Caution

The Median Filter is non-linear — it cannot be described as a simple convolution. This means it does not combine cleanly with frequency-domain analysis. It also tends to erode thin lines or small features if the kernel is too large. Use kernel size 3×3 as a starting point.

4. Bilateral Filter

Non-Linear Spatial
Bilateral filtering is an edge-preserving smoothing technique used in image processing, meaning — reduce noise while maintaining sharp boundaries in an image.

How it works?

Bilateral Filter are based on two factors:

Spatial distance: Like a regular Gaussian Filter, nearby pixels have higher influence on the output.
Intensity similarity: Pixels with similar intensity (brightness) contribute more.

Bilateral Filter combines both, pixels that are spatially close AND have a similar intensity to the center pixel are blended together $⟹$ pixels across an edge and large intensity difference are largely ignored. This ensures that only similar pixels are blended, while edges are preserved.

Formula

I_{f i l t e r e d} (x) = \frac{1}{W} \sum_{x_{i} \in Ω} I (x_{i}) \cdot f_{r} (∥ I (x_{i}) - I (x) ∥) \cdot g_{s} (∥ x_{i} - x ∥)

The Core Components

$I_{filtered} (x)$ : The new, filtered intensity value for the current pixel at position $x$ .
$Ω$ : The local neighborhood (window) of pixels around the target pixel $x$ .
$I (x_{i})$ : The original intensity of a neighboring pixel at position $x_{i}$ .
$\frac{1}{W}$ : A normalization factor that ensures the weights sum up to 1, preventing the image from becoming unnaturally brighter or darker.
The Two Weights
$g_{s} (| x_{i} - x |)$ : Spatial Kernel, this is a standard Gaussian function. If the neighboring pixel $x_{i}$ is physically close to the target pixel $x$ , the weight is high. As you move further away in the window, the weight drops toward zero.
$f_{r} (| I (x_{i}) - I (x) |)$ : Intensity (range) Kernel, This is the edge-preserving element. It looks at the difference in brightness/color between the neighbor and the target. If the neighbor has a very similar intensity to the target, the weight is high. If there is a massive jump in intensity, the weight drops to near zero.

Key Features

mindmap
  root((Bilateral
Filter))    
    Edge Preservation
    Edge-Aware Smoothing
    Dual Gaussian Filters
      Spatial Gaussian 
      Intensity Guassian
    Parameter Sensitivity
      Spatial
\sigma_s
      Intensity
\sigma_r
    Noise Reduction
    Computational
Complexity
        Gaussian
Spatial + 
        Gaussian
Intensity

When to use

Preserves edges excellently: When you need to denoise but must keep edges sharp (e.g., portrait smoothing, medical imaging).
As a preprocessing step before edge detection on natural images. (Works well for natural images where boundaries are important)
Produces visually sharper results compared to Gaussian blur
Performs smoothing by adjusting the weight of each pixel based not only on its spatial distance but also its intensity difference from the target pixel.

Parameters to tune

$σ_{s}$ (spatial): controls how far spatially neighbors influence the center. Larger = more spatial blending.
$σ_{r}$ (range/intensity): controls how different in intensity a pixel can be and still contribute. Smaller = stronger edge preservation.

When to avoid

Real-time or embedded systems — the Bilateral Filter is significantly slower than gaussian filter.
Very noisy images with large noise amplitudes — the range kernel may fail to distinguish noise from true edges.
When edge sharpness is not a concern — Gaussian is simpler and faster.

Disadvantages

Computationally Intensive: Bilateral Filter needs computing both, the intensity and spatial weights for each pixel ( $\approx 2 x$ ) makes the bilateral filter slower than other smoothing filters like Mean and Gaussian filter, especially for large images.
Parameter Tuning: In this filter the parameter tuning for choosing appropriate values for the spatial and intensity sigma values can be tricky and requires experimentation for different images.

Caution

Bilateral filtering can create a "cartoon" or "plastic" appearance if over-applied, known as the HDR effect. It can also fail to separate noise from real edges when noise amplitude is large relative to the actual edge contrast.

5. Min & Max Filters

Non-Linear Spatial

Smoothing & Blurring Filters

Classification of Smoothing Filters

1. Mean (Average) Filter

When to use ?

When to avoid?

2. Gaussian Filter

How it works?

The 2D Gaussian Formula

Step-by-Step Example

1. Create the Kernel Weights

2. Normalize the Weights

3. Apply the Filter (Convolution)

4. Result

When to use ?

How to pick σ ?

When to avoid ?

3. Median Filter

How it works?

When to use?

When to avoid

4. Bilateral Filter

How it works?

Formula

Key Features

When to use

Parameters to tune

When to avoid

Disadvantages

5. Min & Max Filters

How to pick $σ$ ?