RV COLLEGE OF ENGINEERING ®,BENGALURU-560059

(Autonomous Institution Affiliated to VTU, Belagavi)

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING

Computer Vision EL 2 - Concept Explanation Activity

Submitted by,

Name USN Topic

Namrata Srinivasa 1RV23CS146 Wiener Filtering

Navami Lokesh 1RV23CS147 Geometric Mean Filter

Navya Madiraju 1RV23CS148 Inverse Filtering

Nayana Jagadish Raikar 1RV23CS149 Constrained Least Square

Filtering

in partial fulfillment for the requirement of 6th Semester

COMPUTER VISION (CS365TDA)

Under the Guidance of

Prof. Mekhala Vinod Purohit,
Assistant Professor,
Computer Science and Engineering,
RV College of Engineering
Academic Year 2025 - 2026

1
 Inverse Filtering in Digital Image Processing

1. Background
In digital image processing, the goal of image restoration is to recover an original image from a
degraded observation. In practical scenarios, images rarely remain perfect during acquisition,
transmission, or storage. Various physical and environmental factors such as motion during capture,
defocus in optical systems, atmospheric turbulence, and sensor noise introduce distortions into the
image. As a result, the observed image differs significantly from the true scene.

This degradation process is commonly modeled using the following equation:

g(x,y)=h(x,y)∗f(x,y)+n(x,y)

Here, ( f(x,y) ) represents the original image, ( h(x,y) ) denotes the degradation function or point spread
function, ( n(x,y) ) represents additive noise, and ( g(x,y) ) is the observed degraded image. This model
clearly indicates that the observed image is not a direct representation of the original scene, but rather a
blurred and noisy version of it. The central challenge in image restoration lies in estimating the original
image from this imperfect observation, which is often difficult because some information may already
be lost during degradation.

2. Convolution and the Nature of Image Degradation

A key component of the degradation model is convolution, which mathematically describes how the
original image is transformed into a blurred version. Convolution is an operation in which each pixel in
the output image is influenced by a weighted combination of its neighboring pixels in the input image.
This means that pixel intensities are no longer independent; instead, they spread across nearby regions.

The degradation function ( h(x,y) ), also known as the point spread function, characterizes how a single
point in the image is distributed spatially. For instance, in motion blur, the intensity of a point spreads
along the direction of motion, whereas in defocus blur, it spreads outward in a circular pattern. In
atmospheric turbulence, the spreading is more random and irregular.

From a physical perspective, convolution models how imperfections in imaging systems cause the loss
of sharpness. Fine details and edges, which correspond to rapid changes in intensity, are smoothed out

2
due to this spreading effect. Therefore, convolution is the fundamental reason why images appear
blurred in real world scenarios, and any restoration technique must attempt to reverse this process.

3. Frequency Domain Representation

Solving the degradation model directly in the spatial domain is often mathematically complex due to the
convolution operation. To simplify the problem, the image is transformed into the frequency domain
using the Fourier Transform. This transformation represents the image as a combination of sinusoidal
components of varying frequencies.

In the frequency domain, the degradation model becomes:

G(u,v)=H(u,v)⋅F(u,v)+N(u,v)

In this representation, ( F(u,v) ), ( H(u,v) ), and ( G(u,v) ) are the Fourier transforms of the original
image, degradation function, and observed image respectively, while ( N(u,v) ) represents noise in the
frequency domain.

The major advantage of this transformation lies in the fact that convolution in the spatial domain is
converted into multiplication in the frequency domain. This significantly simplifies the mathematical
treatment of the problem. Additionally, analyzing images in terms of frequency provides deeper insight
into their structure. Low-frequency components correspond to smooth regions of the image, while
high-frequency components correspond to edges and fine details. Since blur primarily reduces
high-frequency content, understanding this representation is essential for restoration.

4. Fast Fourier Transform (FFT)

Although the Fourier Transform provides a powerful framework for analyzing images, its direct
computation is computationally expensive, especially for large images. To address this, the Fast Fourier
Transform (FFT) is used as an efficient algorithm for computing the Discrete Fourier Transform.

The FFT reduces the computational complexity from ( O(N^2) ) to ( O(N \log N) ), making it feasible to
perform frequency domain operations in practical applications. This improvement is crucial because
modern image processing systems deal with large datasets and often require real-time performance. As a
result, FFT is widely used in almost all frequency-based image processing techniques, including inverse
filtering.

3
5. Concept of Inverse Filtering
Inverse filtering is one of the simplest techniques used for image restoration. It is based on the idea of
reversing the degradation process by applying the inverse of the degradation function in the frequency
domain.

From the frequency domain degradation model:

G(u,v)=H(u,v)⋅F(u,v)

the original image can be estimated as:

F(u,v)=H(u,v) / G(u,v)​

This equation represents the core principle of inverse filtering. By dividing the degraded image spectrum
by the degradation function, the effect of blurring can theoretically be removed. The restored image is
then obtained by applying the inverse Fourier Transform:

f^​(x,y)=F^-1 {F(u,v)}

Conceptually, inverse filtering attempts to “undo” the spreading effect caused by convolution. If the
degradation function is perfectly known and no noise is present, this method can recover the original
image exactly.

6. Procedure of Inverse Filtering

The inverse filtering process follows a structured sequence. First, the degraded image is obtained and
transformed into the frequency domain using the Fast Fourier Transform. The degradation function is
also represented in the frequency domain. The core operation involves dividing the transformed
degraded image by the degradation function for each frequency component. Finally, the inverse Fourier
Transform is applied to reconstruct the restored image in the spatial domain.

This pipeline highlights how the transformation to the frequency domain simplifies the restoration
process, allowing complex convolution operations to be replaced with straightforward division.

4
7. Assumptions and Practical Challenges
Despite its simplicity, inverse filtering relies on strong assumptions. It assumes that the degradation
function is known exactly and that the noise component is negligible. In practical scenarios, these
assumptions rarely hold true. The exact form of the degradation function is often unknown or difficult to
estimate accurately, and noise is almost always present in real-world images.

Because of these unrealistic assumptions, inverse filtering becomes highly sensitive to errors and often
produces unstable results. This limits its direct applicability in practical image restoration tasks.

8. Limitations of Inverse Filtering

One of the most significant limitations of inverse filtering is noise amplification. When the
degradation function has very small values at certain frequencies, dividing by these values
causes the noise component to be amplified disproportionately. As a result, the restored image
may become noisier than the original degraded image.

Another major issue is the problem of division by zero. If the degradation function becomes zero
at any frequency, the division operation becomes undefined, leading to irreversible loss of
information. This means that certain components of the original image cannot be recovered
under any circumstances.

Inverse filtering is also highly unstable, as small errors in estimating the degradation function or
noise can lead to large deviations in the restored image. Additionally, since blur suppresses
high-frequency components and noise is often dominant in those regions, inverse filtering tends
to enhance noise while attempting to restore details.

9. Practical Considerations and Extensions

Due to the limitations of inverse filtering, more advanced restoration techniques have been developed.
Methods such as Wiener filtering incorporate statistical information about noise and signal to achieve a
balance between restoration and noise suppression. Similarly, constrained least squares filtering
introduces regularization to control instability and improve robustness.

5
These methods can be viewed as extensions of inverse filtering, addressing its shortcomings while
retaining its fundamental principles.

10. Conclusion
Inverse filtering is a fundamental technique in digital image processing that provides a straightforward
approach to image restoration by reversing the degradation process in the frequency domain. While it
offers valuable theoretical insight and forms the basis for more advanced methods, its practical use is
limited due to sensitivity to noise, instability, and the requirement of precise knowledge of the
degradation function. Consequently, inverse filtering is primarily used as a conceptual foundation rather
than a standalone solution in real-world applications.

6
 Wiener Filtering
1. Background

In image restoration, the main goal is to recover the original image from a degraded version that has
been affected by blur and noise. This degradation is commonly modeled as:

This equation essentially states that the image we observe is not the true image, but rather a blurred
version of it with noise superimposed. Therefore, the restoration problem is inherently challenging
because we are trying to reverse two effects simultaneously: blur and noise.

A straightforward approach such as inverse filtering attempts to directly undo the blur by dividing by the
PSF in the frequency domain. However, this method is highly unstable in practice. Whenever the PSF
has very small values, the division leads to extreme amplification of noise, resulting in a highly distorted
output.

The Wiener filter addresses this limitation by introducing a more balanced and realistic approach.
Instead of blindly inverting the degradation, it incorporates knowledge about the signal and noise to
produce a stable and meaningful estimate.

7
2. Core Idea of Wiener Filtering

The Wiener filter is based on a statistical formulation of the restoration problem. Rather than enforcing
constraints like smoothness (as in CLS filtering), it seeks to minimize the average error between the
restored image and the true image. This is expressed as:

This formulation means that the Wiener filter does not aim for exact reconstruction in every pixel.
Instead, it aims to produce the best possible estimate on average, taking into account the randomness of
noise.

In simpler terms, the filter tries to answer the question: given what we observe and what we know about
noise, what is the most reliable estimate of the original image? This makes the Wiener filter
fundamentally different from deterministic approaches.

8
3. Frequency Domain Representation

To simplify both the mathematical analysis and practical implementation, the degradation model is
typically transformed into the frequency domain using the Fourier transform. In this domain,
convolution becomes simple multiplication, which makes the restoration process much more tractable.
The degradation model can then be written as:

This representation is particularly useful because it allows us to analyze how different frequency
components of the image are affected by blur and noise. Typically, blur attenuates high-frequency
components, while noise is often spread across all frequencies.

The restoration process involves applying a filter in the frequency domain to recover an estimate of the
original image:

This formulation highlights an important idea: instead of directly modifying pixel values in the spatial
domain, we operate on frequency components. Each frequency is treated independently, allowing the
filter to selectively enhance or suppress specific components depending on their reliability. This
frequency-wise control is what makes Wiener filtering both flexible and effective.

9
4. Wiener Filter Expression

The Wiener filter is mathematically defined as:

This equation provides a clear insight into how Wiener filtering operates. The numerator, which contains
the complex conjugate of the PSF, attempts to reverse the blurring process. If this term existed alone, the
filter would behave like an inverse filter. However, the denominator introduces a crucial stabilizing
𝑆𝑛
factor. The term 𝑆𝑢
represents the ratio of noise power to signal power at each frequency. This means

that the filter automatically accounts for how reliable each frequency component is. If a particular
frequency is heavily corrupted by noise, the denominator becomes large, reducing the filter response at
that frequency.

This adaptive balancing between inversion and suppression is what makes Wiener filtering robust.
Unlike simple inversion, it avoids amplifying unreliable components, leading to more stable and visually
meaningful restorations.

5. Intuition Behind the Filter

The Wiener filter can be better understood by examining its behavior under different noise conditions.
Its response is not uniform across frequencies; instead, it adapts based on the relative strength of signal
and noise. When the noise is negligible compared to the signal, the Wiener filter behaves approximately
like an inverse filter:

10
In this case, the filter focuses on restoring sharp details by undoing the blur. However, when noise is
significant, the filter suppresses those frequencies and 𝐺(ω) ≈ 0. This adaptive behavior is the key
strength of Wiener filtering. Instead of treating all frequencies equally, it selectively restores or
suppresses components based on their reliability.

6. Practical Form (Simplified Wiener Filter)

In real-world applications, the exact power spectral densities of the signal and noise are rarely known.
Estimating them accurately is often difficult and computationally expensive. Therefore, a simplified
version of the Wiener filter is commonly used:

In this formulation, the term K serves as a simplified representation of the ratio between noise power
and signal power. Although it does not capture frequency-dependent variations, it still provides a useful
way to control the behavior of the filter. The parameter K plays a role similar to the regularization
parameter in CLS filtering. A small value of K results in a filter that closely resembles inverse filtering,
producing sharper images but potentially amplifying noise. Conversely, a large value of K leads to
stronger noise suppression at the cost of reduced detail and increased smoothness.

11
In practice, K is often chosen empirically. One common approach is to start with a small value and
gradually increase it until a visually acceptable balance between noise reduction and detail preservation
is achieved. Despite its simplicity, this approximation works well in many practical scenarios and is
widely used in image processing applications.

7. Matrix Interpretation

The degradation model can also be expressed in vector form:

The Wiener solution in matrix form is given by:

This formulation highlights that Wiener filtering is fundamentally a statistical estimation technique. It
incorporates prior knowledge about both the signal and the noise, making it more robust compared to
purely deterministic methods.

12
8. Summary

Wiener filtering provides a powerful and theoretically grounded approach to image restoration. By
minimizing the mean square error, it produces an optimal estimate of the original image under given
statistical assumptions.

Unlike inverse filtering, it avoids instability, and unlike CLS filtering, it does not rely on manually
imposed smoothness constraints. Instead, it adapts automatically based on the relative strength of signal
and noise across frequencies.

Overall, Wiener filtering achieves a careful balance between restoring image details and suppressing
noise, making it one of the most widely used and effective techniques in practical image restoration
problems.

13
Constrained Least Squares (CLS) Filtering
1. Background
In image restoration, the core goal is to recover an original (ideal) image from a degraded observation.
The observation is typically modeled as:

v(m,n) = h(m,n) ⊗ u(m,n) + n(m,n)

Where,

1.​ h(m,n) : is the point spread function (PSF) representing the degradation or blur
2.​ u(m,n) : is the unknown ideal image
3.​ n(m,n) : is additive noise
4.​ v(m,n) : is the observed or degraded image.

This equation basically says that what we observe is not the original image, but a blurred version of it
with some noise added on top. So the task is not just to remove blur, but to do it carefully without
boosting the noise. That’s where things get tricky, because directly reversing the blur tends to blow up
the noise.

The simplest inversion such as the inverse filter simply divides by the PSF in the frequency domain. But
this is practically useless because it amplifies noise catastrophically. Better alternatives include the
Wiener filter (which assumes knowledge of power spectral densities) and the family of parametric
estimation filters. The constrained least squares (CLS) filter is a powerful member of this family: it does
not require prior statistical knowledge yet still controls noise amplification by penalizing solution
roughness.

2. The Constrained Least Squares Problem

2.1 Core Idea

The CLS filter produces an estimate û(m,n) of the ideal image by minimizing a smoothness criterion,
specifically the energy in a high pass filtered version of the estimate, subject to a data fidelity constraint.
Formally:

Minimize J ≜ ‖q(m,n) ⊗ û(m,n)‖²

14
 Subject to ‖v(m,n) − h(m,n) ⊗ û(m,n)‖² ≤ ε²

Here,

1.​ q(m,n): is a roughness measuring operator (e.g., Laplacian or high pass filter)
2.​ ε² ≥ 0 : is the tolerated residual energy which is linked to noise level

So the idea behind CLS is pretty intuitive when you think about it. We want the restored image to still
resemble the observed image, but at the same time we don’t want it to be too noisy or full of sudden
intensity changes. So we add a smoothness condition. In simple terms, we are trying to find an image
that fits the data reasonably well but is also not “too rough”.

The objective J penalizes oscillations or the roughness in the estimate, while the constraint ensures the
restored image remains consistent with the observed data.

2.2 Parseval's Theorem

Applying Parseval's theorem, the spatial domain norms become frequency domain integrals:

‖a(m,n)‖² = (1/4π²) ∬ |A(ω₁,ω₂)|² dω₁dω₂ ≜ ‖A(ω₁,ω₂)‖²

This transforms the constrained minimization into frequency domain, making it tractable via the
Lagrange multiplier method. The objective and constraint become:

Minimize J = ‖Q(ω₁,ω₂) Û(ω₁,ω₂)‖²

Subject to ‖V(ω₁,ω₂) − H(ω₁,ω₂)Û(ω₁,ω₂)‖² ≤ ε²

We move everything to the frequency domain mainly because it simplifies life. Convolution turns into
multiplication, and the optimization becomes much easier to handle mathematically. Most restoration
filters are easier to understand and design in this domain anyway.

3. Deriving the CLS Transfer Function

3.1 Lagrange Multiplier Solution

Using the Lagrange multiplier method, the solution that minimizes J subject to the constraint takes the
form:

Û(ω₁,ω₂) = Gᵧ(ω₁,ω₂) · V(ω₁,ω₂)

15
From this expression, the numerator tries to undo the blur using the conjugate of the PSF. The
denominator prevents instability by avoiding division by very small values of |H|². The additional term
γ|Q|² plays an important role in suppressing high-frequency noise.

where the CLS filter transfer function Gᵧ is:

Gᵧ ≜ H*(ω₁,ω₂) / [|H(ω₁,ω₂)|² + γ|Q(ω₁,ω₂)|²]

The scalar parameter γ (gamma) is the Lagrange multiplier, which controls the trade off between fidelity
to observations and smoothness of the solution.

The parameter γ is very important because it controls the trade off between sharpness and smoothness. If
γ is very small, the filter behaves like an inverse filter, giving a sharper but noisy result. If γ is large, the
output becomes smoother, but fine details may be lost. So γ acts as a tuning parameter.

This is structurally identical to the Wiener filter, but instead of signal or noise power spectra (which are
hard to know), the CLS filter uses γ|Q|² as a tunable regularization term.

3.2 Determining the Lagrange Multiplier γ

The multiplier γ is determined by enforcing the constraint at equality (active constraint). Substituting
and into constraint and requiring equality yields the nonlinear equation in γ:

f(γ) ≜ (γ²/4π²) ∬ [|Q|⁴|V|² / (|H|² + γ|Q|²)²] dω₁dω₂ − ε² = 0

This must generally be solved numerically (e.g., bisection or Newton's method). In practice, if noise
variance σ² is known, one sets ε² = N·σ² (where N is the image size).

4. The Roughness Operator q(m,n)

4.1 Laplacian as Roughness Measure

The most common choice for q(m,n) is a finite difference approximation of the 2D Laplacian ∂²/∂x² +
∂²/∂y². On a square grid with spacing Δx = Δy = 1 and α = ¼:

q(m,n) = −δ(m,n) + α[δ(m−1,n) + δ(m+1,n) + δ(m,n−1) + δ(m,n+1)]

Taking the Fourier transform gives the frequency domain roughness operator:

Q(ω₁,ω₂) = −1 + 2α cos ω₁ + 2α cos ω₂

16
Since Q acts as a high pass filter, ‖Q⊗û‖² is large when û has many high frequency components (sharp
edges, noise), and small when û is smooth. Minimizing J thus favors smooth estimates of exactly what
we want for denoising.

|Q(ω)|² ≈ 0 at low frequencies i.e. smooth regions and large at high frequencies. So, the regularization
term γ|Q|² mainly suppresses high frequency noise without much affecting coarse image structure.

The Laplacian operator emphasizes rapid intensity changes, which correspond to edges and high
frequency components. Since noise also lies in the high frequency region, minimizing this term helps in
reducing noise and producing a smoother image.

4.2 Effect on the Filter

In regions of the frequency plane where:

1.​ The blur |H|² is large (passband of the PSF): γ|Q|² is negligible, so Gᵧ ≈ H*/|H|² = 1/H . Hence it
behaves like an inverse filter.
2.​ The blur |H|² is small (stopband of PSF): γ|Q|² dominates. Thus it suppresses noise amplification.

This graceful frequency dependent behaviour is the fundamental advantage of the CLS filter.

This shows that the CLS filter behaves differently across frequencies. In regions where the blur is small,
it acts like an inverse filter and restores details. In regions where the blur is severe, it suppresses noise
instead of amplifying it. This adaptive behavior is a key advantage.

5. Vector (Matrix) Formulation

5.1 Vector(Matrix) Model

For a spatially invariant PSF, the image observation model can be written in vector form as:

o = H·u + n

Where,

1.​ o - observed
2.​ u - ideal
3.​ n - noise are column vectors of length N² (lexicographically ordered pixels)
4.​ H - is the M×N PSF convolution matrix.

17
The constrained minimization problem in matrix form is:

min ‖ℒ û‖² subject to ‖o − H û‖² ≤ ε²

where, ℒ is the matrix version of the roughness operator q(m,n).

5.2 Matrix Solution

The Lagrange solution gives the closed-form matrix estimate:

û = (γℒᵀℒ + Hᵀ H)⁻¹ Hᵀ o

and γ satisfies the constraint equation:

γ²‖[H(ℒᵀℒ)⁻¹Hᵀ + I]⁻¹o‖² − ε² = 0

In the absence of blur (H = I), the matrix solution collapses to the smoothing spline:

û = (γℒᵀℒ + I)⁻¹ o

This is the 2D discrete version of the smoothing spline, showing that spline smoothing is a special case
of CLS filtering when there is no blur.

In this form, the term involving H ensures that the solution remains consistent with the observed data,
while the term involving ℒ enforces smoothness. The final solution is therefore a balance between data
fidelity and regularization.

6. Relationship to Other Filters

6.1 CLS vs. Wiener Filter

Property Wiener Filter CLS Filter

Requires power Yes — Suu and Snn needed No , only ε² (noise energy)
spectra?

Regularization term Wn / Wu (noise to signal ratio) γ|Q(ω)|² (roughness penalty)

Optimality criterion Minimum mean square error Min roughness, constrained fidelity

Noise model Yes (statistical model) No (just energy bound ε²)

assumed?

Parameter to tune None (if spectra known) γ (Lagrange multiplier)

18
6.2 CLS and the Constrained Least Squares Filter of Hunt

Another parametric variant, the constrained least squares filter of Hunt, is given by:

H_R(ωx,ωy) = H*_D(ωx,ωy) / [|H_D(ωx,ωy)|² + γ|C(ωx,ωy)|²]

Where,

1.​ γ is a design constant

2.​ C(ωx,ωy) is a design spectral variable.

6.3 CLS as a Generalized Inverse / Pseudoinverse

When there is no regularization (γ → 0), the matrix solution û = (HᵀH)⁻¹Hᵀo is exactly the pseudoinverse
(least norm least squares) solution. The pseudoinverse H⁻ = (HᵀH)⁻¹Hᵀ satisfies H⁻H = I, making CLS a
regularized pseudoinverse.

7. Implementation Steps (Frequency Domain)

1.​ Compute the DFT of the observed image: V(ω₁,ω₂) = DFT{v(m,n)}

2.​ Compute the DFT of the PSF: H(ω₁,ω₂) = DFT{h(m,n)} (zero pad to image size)

3.​ Compute the DFT of the roughness operator: Q(ω₁,ω₂) = DFT{q(m,n)} (e.g., Laplacian)

4.​ Choose ε² ≈ N·σ² where σ² is the noise variance, or estimate from a flat image region.

5.​ Solve f(γ) = 0 numerically for γ (e.g., bisection search).

6.​ Compute CLS filter: Gᵧ = H* / (|H|² + γ|Q|²)

7.​ Compute estimate: Û = Gᵧ · V

8.​ Reconstruct: û(m,n) = IDFT{Û(ω₁,ω₂)} (take real part)

When ε² is unknown, start with a small γ (less smoothing) and increase until the restored image looks
visually acceptable. This is equivalent to regularization parameter selection in Tikhonov regularization.
In practice, choosing the value of γ is crucial. If the noise level is known, it can be set accordingly.
Otherwise, γ is adjusted experimentally by starting with a small value and increasing it until a good
balance between noise reduction and detail preservation is achieved.

8. Key Properties and Observations

8.1 Stability
19
Unlike the inverse filter, CLS never has division by zero issues because |H|² + γ|Q|² > 0 for all
frequencies (as long as γ > 0 and Q ≠ 0 everywhere, which holds for the Laplacian).

8.2 Smoothing Splines as a Special Case

When H = I (no blur and pure denoising), the CLS frequency domain filter becomes Gᵧ = 1/(1 + γ|Q|²),
which is exactly the smoothing spline. The roughness penalty γ|Q|² cuts out high frequency noise while
retaining low frequency image content.

8.3 Connection to Tikhonov Regularization

CLS filtering is equivalent to Tikhonov regularization with regularization matrix ℒ. The constrained
minimization form (min roughness s.t. fidelity constraint) is dual to the unconstrained Tikhonov form:
min{‖v − Hû‖² + γ‖ℒû‖²}, with the same solution.

8.4 Limitation

CLS produces a single linear filter for the entire image. It cannot adapt locally, regions with edges are
over smoothed while flat regions may retain noise. Spatially adaptive variants and iterative methods
(Landweber, conjugate gradient) are needed for better performance in such cases.

Overall, the CLS filter provides a practical way to restore images affected by blur and noise without
requiring detailed statistical information. It avoids the instability of inverse filtering and does not depend
on power spectra like the Wiener filter. Instead, it uses a smoothness constraint controlled by γ, allowing
it to preserve important image structures while reducing noise.

20
 GEOMETRIC MEAN FILTERING

1. Introduction to Image Restoration

Image restoration is an important area of digital image processing that focuses on reconstructing an
image that has been degraded during acquisition, transmission, or storage. In real-world applications,
images are often affected by various types of noise and distortions such as sensor noise, motion blur,
atmospheric disturbances, and transmission errors. These degradations reduce the quality and usability
of the image.

The primary objective of image restoration is to recover the original image as accurately as possible
using mathematical models and filtering techniques. Among the various restoration methods, spatial
domain filtering techniques are widely used due to their simplicity and effectiveness. One such
important category is the mean filter family, which includes the geometric mean filter.

2. Mean Filter Family

Mean filters are a class of spatial domain filters that operate by replacing each pixel value with a value
computed from its neighboring pixels. These filters are primarily used for smoothing images and
reducing noise. The different types of mean filters include arithmetic mean, geometric mean, harmonic
mean, and contraharmonic mean filters.

While the arithmetic mean filter computes the simple average of pixel values, it often leads to loss of
image details and excessive blurring. In contrast, the geometric mean filter uses a multiplicative
approach, which helps in preserving image details while still achieving effective noise reduction. This
makes it more suitable for applications where maintaining image sharpness is important.

3. Geometric Mean Filtering

Geometric mean filtering is a nonlinear filtering technique used for image restoration.

3.1 Key Characteristics

1.​ Operates in the spatial domain

2.​ Uses multiplication instead of addition
21
 3.​ Processes pixels within a local neighborhood
4.​ Preserves edges better than arithmetic mean filter

4. Definition

For a corrupted image ( g(x, y) ), the geometric mean filter computes the restored image ( \hat{f}(x, y) )
using the geometric mean of pixel values within a neighborhood.

4.1 Concept

1.​ Multiply all pixel values

2.​ Take the ( mn )-th root
3.​ Replace the center pixel

5. Mathematical Formulation

The geometric mean filter is expressed as:

5.1 Parameters

1.​ (f^(x,y) ): Restored pixel

2.​ ( g(r,c) ): Corrupted image pixel values
3.​ ( Sxy ): Neighborhood window
4.​ ( mxn ): Window size
5.​ ( mn ): Total pixels
6.​

5.2 Interpretation

The formula computes the geometric mean, reducing the influence of extreme values.

22
6. Working Principle

1.​ The geometric mean filter operates using a sliding window approach across the image. A
window of size m×nm \times nm×n, typically with odd dimensions such as 3×3 or 5×5, is
selected and centered on each pixel of the image. All pixel values within this window are
multiplied together, and the resulting product is raised to the power 1/mn1/mn1/mn to compute
the geometric mean.
2.​ The computed value replaces the original pixel at the center of the window. This process is
repeated for every pixel in the image by moving the window systematically across the entire
image. Through this method, noise is reduced while preserving important image features.

7. Geometric Mean Concept

The geometric mean of ( n ) numbers is defined as:

7.1 Example (3×3 Window)

1.​ Total pixels = 9

2.​ Multiply all pixel values
3.​ Take 9th root

7.2 Comparison with Arithmetic Mean

1.​ Arithmetic mean is more affected by extreme values

2.​ Geometric mean is more stable and preserves details

8. Noise Handling Capability

8.1 Gaussian Noise

1.​ Most common noise type

2.​ Caused by electronic disturbances
23
 3.​ Effectively reduced by geometric mean filter

8.2 Uniform Noise

1.​ Equal probability distribution

2.​ Also handled efficiently

8.3 Additive Noise

1.​ General noise added to image

2.​ Reduced through smoothing

8.4 Limitations in Noise Handling

1.​ Not suitable for impulse (salt-and-pepper) noise

2.​ Median filter is preferred in such cases

9. Comparison with Arithmetic Mean Filter

Feature Arithmetic Mean Filter Geometric Mean Filter

Operation Addition & division Multiplication & root

Noise Reduction Good Comparable

Detail Preservation Moderate High

Edge Sharpness Low Better

Sensitivity to Outliers High Low

Complexity Low Moderate

24
9.1 Key Observation

The geometric mean filter provides similar smoothing but retains more image details.

10. Advantages

One of the major advantages of the geometric mean filter is its ability to preserve edges and fine details
in the image. Unlike arithmetic mean filtering, which can significantly blur the image, the geometric
mean filter maintains sharper transitions between regions. It is also effective in reducing Gaussian and
uniform noise, making it suitable for a wide range of applications.

Additionally, the filter is conceptually simple and can be implemented using a straightforward
neighborhood processing approach. Its flexibility also allows it to be extended into more advanced
filtering techniques.

11. Limitations

Despite its advantages, the geometric mean filter has certain limitations. It is not effective in handling
impulse noise, as the presence of zero-valued pixels can severely affect the output. The filter is also
computationally more complex due to the need for multiple multiplications and exponentiation
operations.

Furthermore, although it preserves details better than arithmetic mean filtering, it still introduces some
degree of blurring. The performance of the filter is also sensitive to the choice of window size, as larger
windows may lead to excessive smoothing.

25
12. Generalized Geometric Mean Filter

The filter can be extended to the frequency domain.

12.1 Features

1.​ Combines inverse filtering and Wiener filtering

2.​ Uses control parameters

12.2 Special Cases

1.​ ( alpha = 1 ): Inverse Filter

2.​ ( alpha = 0 ): Wiener Filter
3.​ ( alpha = 0.5 ): Geometric Mean behavior

12.3 Significance

Provides a unified framework for multiple restoration techniques.

13. Applications

Geometric mean filtering is widely used in various fields of image processing. In medical imaging, it is
used to enhance X-ray and MRI images by reducing noise while preserving important diagnostic details.
In industrial inspection, it helps in detecting defects in circuit boards and manufactured components.

The filter is also used in satellite and aerial imaging to remove noise caused by atmospheric
disturbances. Additionally, it is applied in general image processing tasks where noise reduction is
required without significant loss of detail.

26
14. Conclusion

Geometric mean filtering is an important technique in digital image processing that effectively balances
noise reduction and detail preservation. It performs better than arithmetic mean filtering in maintaining
image quality, especially in the presence of Gaussian noise.

However, its computational complexity and limitations in handling impulse noise must be considered
when applying it in practical scenarios.

27