Digital Image Processing (MAKAUT –

CSE)
1. Introduction

Digital Image Processing (DIP) is the use of computer algorithms to process images
as digital (discrete) data. An image is formally a two-dimensional function f(x,y)
giving intensity at spatial location (x,y). When x, y and f are all finite and discrete, the
image is digital [Link] . Each elemental picture unit (pixel) has a unique
location and intensity value [Link] . DIP originated in the 1960s (e.g. NASA’s
lunar imagery) to enhance and analyze images by computer [Link] [Link] .
Unlike analog processing, DIP can apply many sophisticated algorithms and avoid
cumulative noise [Link] .

A digital image processing system has five main elements [Link] :

Image Acquisition: sensing and sampling (e.g. camera or scanner) to obtain an

analog image.

Image Storage: memory/disk to hold images (raw or compressed).

Image Processing: computational hardware/software to apply algorithms.

Image Communication: transmission (e.g. networking, multiplexing).

Image Display: output devices (monitors, printers).

These elements cover the imaging pipeline from capture through processing to final
display [Link] .

Fundamental Steps in DIP

A typical DIP workflow proceeds in stages [Link] [Link] :

Image Acquisition: Capture analog signals (light, X-ray, etc.) and digitize via
sampling and quantization.
 Enhancement & Restoration: Improve image appearance (contrast, sharpness)
and correct degradations (blur, noise).

Compression: Reduce data size for storage/transmission (e.g. JPEG).

Segmentation: Partition the image into regions (objects/background) by edge

detection, thresholding, or region methods.

Feature Extraction & Recognition: Derive attributes (shapes, textures) and

classify objects (face recognition, OCR).

Early stages (“low-level processing”) take an image in and produce another image;
later stages (“high-level”) extract information or semantics [Link] [Link] .
For example, enhancement and restoration improve visual quality, while segmentation
and recognition enable object detection. As one survey notes, “Image enhancement
improves visual quality (e.g. noise removal via filtering); image restoration uses
knowledge of degradation to undo blur/noise” [Link] [Link] .

2. Digital Image Formation

2.1 A Simple Image Model

In the real world, an image represents light intensity. A common model separates
illumination (amount of incident light i(x,y)) and reflectance (fraction reflected r(x,y)),
so that the observed intensity is

f (x, y) = i(x, y) ⋅ r(x, y) ,

where 0 < i(x, y) < ∞ and 0 < r(x, y) ≤ 1 [Link] . For example, a bright day
(high i) or a shiny object (high r) increases f(x,y). In a monochrome (grayscale) image,
f(x,y) is usually mapped to a gray level l = f (x, y) in a limited range [Lmin , Lmax ], ​ ​

often normalized to [0, L − 1] (with 0 = black, L − 1 = white) [Link] [Link] .

To create a digital image, the continuous intensity function f (x, y) is sampled on a

grid and each sample is quantized to an integer gray level [Link] [Link] . Sampling
discretizes the spatial coordinates: the image plane is overlaid by a regular grid and
f (x, y) is measured at the center of each grid cell [Link] . Mathematically, a digital
image can be viewed as a function f [m, n] where indices m, n are integers (pixel
coordinates) and f [m, n] is the gray value [Link] . Quantization discretizes the
amplitude: each sample value is rounded to one of a finite set of levels. In practice,
common sensors use 8–16 bits per pixel, giving 256–65536 levels [Link] [Link] .
Non-uniform quantization (e.g. μ-law companding) may be used to allocate more
levels where intensities have large perceptual variation.

2.2 Geometric Model – Basic Transformations

Geometric transformations reposition or reshape the image. They include translation,

scaling, and rotation (plus reflection/shear). Using homogeneous coordinates, a 2D
point P = (x, y) can be mapped to P ′ = (x , y )
′ ′
by matrix operations. For
example, a translation by (tx , ty ) is given by
​ ​

′ ′
x = x + tx , ​

y = y + ty , ​

often written as

′
x 1 0 tx ​
x
′
y
​ ​ ​
= 0
​ ​

1 ​

ty ​ ​ ​

y
​ ​ ​
.

1 0 0 1 1

Fig: Geometric translation of a point P (x, y) to P ′ (x + t x , y + t y ). ​ ​

In translation, all
points shift by the same offset (tx , ty ). ​ ​

Scaling changes size by factors Sx , Sy in each dimension: ​ ​

′ ′
x = Sx x, ​
y = Sy y, ​
 Sx 0
implemented by the matrix ( . For example, multiplying coordinates by 2
​

) ​

0 Sy ​

doubles the object’s dimensions [Link] . Rotation by an angle θ about the origin
maps

′ ′
x = x cos θ − y sin θ, y = x sin θ + y cos θ,

cos θ − sin θ
corresponding to the rotation matrix ( ​
)
​
[Link] . Below are
sin θ cos θ

examples of scaling and rotation:

Fig: Scaling an object by factors Sx , Sy (here expanding in one direction and

​ ​

contracting in the other). The image is multiplied by the diagonal matrix diag(Sx , Sy ) ​ ​

[Link] .

Fig: Rotation of point P (x, y) by angle θ about the origin. The coordinates transform by
= x cos θ − y sin θ , y .
′ ′
x = x sin θ + y cos θ [Link]

Such linear transformations apply to each pixel’s coordinates. More complex

transformations (affine or projective) include shearing or perspective effects, but the
basic translation, scaling, and rotation are fundamental building blocks.

2.3 Perspective Projection

Perspective projection models how a 3D scene is imaged by a pinhole or camera. A

scene point (X , Y , Z ) projects to an image point (x, y) according to

X Y
x = f ​

, y = f ​

,
Z Z

where f is the camera focal length [Link] . Equivalently X /Z = x/f , indicating

that farther objects (larger Z ) appear closer to the vanishing point. Perspective
projection causes parallel lines in 3D to converge in the image, and distant objects to
appear smaller. In practice, camera calibration determines f and corrects for lens
distortion; but the pinhole model (x = X f /Z ) is the classical basis [Link] .

2.4 Sampling and Quantization

As noted, sampling converts a continuous image into discrete samples on a grid, and
quantization maps each sample’s intensity into a finite set of values. Uniform
sampling uses a fixed grid spacing (e.g. pixels equally spaced); non-uniform sampling
(e.g. hexagonal or adaptive sampling) uses variable spacing to capture features more
efficiently. Likewise, uniform quantization divides the intensity range into equal-size
levels, while non-uniform quantization uses variable step sizes (often to match
human perception or signal statistics) [Link] [Link] . In practice, image
sensors sample uniformly and apply uniform quantization (8–12 bits) [Link]

[Link] . Non-uniform quantizers (μ-law, A-law) are used in some systems (audio,
telephony) to give higher precision near low amplitudes. Uniform quantization is
simple and optimal for signals with high SNR, whereas non-uniform quantization can
reduce error for signals with non-uniform amplitude distributions [Link] .

3. Mathematical Preliminaries

3.1 Pixel Neighbors and Connectivity

In a digital image, each pixel has adjacent neighbors. For a pixel p = (x, y) , the 4-
neighbors (N4 (p)) are the pixels that share an edge with p: (x
​
± 1, y) or (x, y ± 1)

[Link] . The 8-neighbors (N8 (p)) include the 4-neighbors plus the 4 diagonal
​

neighbors (x ± 1, y ± 1) [Link] . (Often one excludes diagonal neighbors

sharing only a corner in certain connectivity definitions.) Connectivity defines how
pixels join into regions. Two pixels are 4-connected if there is a path between them
moving only among 4-neighbors; similarly, 8-connected allows diagonal moves.
Formally, a set of pixels P is 4-connected if for every pair pi , pj ​ ​
∈ P there exists a
sequence of pixels in P where each consecutive pair is 4-neighbors [Link] . (8-
connected is analogous with 8-neighbors.) These adjacency relations are reflexive
and transitive, so they generate equivalence classes (connected components) under
transitive closure. Connectivity is crucial in morphological and segmentation
algorithms to define regions and boundaries.

3.2 Distance Measures

Common distance metrics between pixel coordinates (x1 , y1 ) and (x2 , y2 ) include: ​ ​ ​ ​

Euclidean distance: DE ​

= (x1 − x2 )
​ ​
2
+ (y1 − y2 )
​ ​
2
.
​

City-block (Manhattan) distance: D1 ​

= ∣x1 − x2 ∣ + ∣y1 − y2 ∣
​ ​ ​ ​

.
 Chessboard (Chebyshev) distance: D∞ ​
= max(∣x1 − x2 ∣, ∣y1 − y2 ∣)
​ ​ ​ ​
.

These measure “closeness” for region-growing or filtering operations. Euclidean is

rotation-invariant; D1 and D∞ are simpler to compute on grids. Weighted or
​ ​

quadratic forms can also be used for anisotropic metrics.

3.3 Arithmetic and Logical Operations

Pixel-wise arithmetic operations are fundamental for combining or modifying images.

Examples include: addition/subtraction of images (e.g. image differencing),
multiplication by constants (brightness adjustment), or pixel-wise multiplication. In
spatial domain, these correspond to point operations. Logical (bitwise) operations
include AND, OR, XOR, which can implement masks and segmentation (e.g.
thresholding via a binary mask). For example, if g(x, y) is a binary mask (0 or 1), then
f (x, y) AND g(x, y) zeroes out pixels outside a region. Logical operations are fast
and often used in pre- or post-processing (e.g. bit-plane slicing).

3.4 Fourier Transform (2D)

Fourier analysis decomposes an image into frequency components. The 2D discrete

Fourier transform (DFT) of an M × N image f [m, n] is

M −1 N −1
um vn
−j 2π( ​

+ ​

)
F (u, v) = ∑ ∑ f (m, n) e ​ ​
M N
.

m=0 n=0

This transforms spatial coordinates (m, n) into frequency coordinates (u, v)

[Link] . The inverse DFT reconstructs the image. The DFT is generally
complex-valued; images typically display the magnitude (spectrum) ∣F (u, v)∣. The
2D DFT has the separability property: it can be computed by first taking 1D DFTs of
each row, then of each column [Link] [Link] .

Key properties of the 2D DFT include linearity, convolution, shift, and periodicity
[Link] [Link] :

Linearity: F {af + bg} = aF + bG for constants a, b [Link] .

Convolution theorem: Circular convolution in space corresponds to

multiplication in frequency, f ∗ g = F
−1
{F ⋅ G} [Link] .
 Separability: The 2D DFT can be done by 1D transforms on rows and columns
(as above) [Link] .

Periodic (Discrete): F (u, v) is periodic in u, v with periods M , N [Link] .

Shift (Translation): A spatial shift by (m0 , n0 ) introduces a linear phase factor in

​ ​

the DFT:
um0 vn0
.
​ ​

F {f [m − m0 , n − n0 ]} = F (u, v) exp[ − j 2π(

​ ​ ​

+ ​

)] [Link]
M N

Rotation: Rotating an image by angle α rotates its spectrum by the same angle
[Link] .

These properties underlie many filtering and restoration methods.

3.5 Discrete Cosine and Sine Transforms

The Discrete Cosine Transform (DCT) is a closely related transform that uses only
real cosine basis functions. The 2D DCT of an image expresses it as a sum of cosine
waves of varying frequency and amplitude. Because it concentrates most signal
energy into a few coefficients, the DCT is widely used in image compression (e.g.
JPEG) [Link] . In fact, “the DCT… is widely used in most digital media, including
digital images (such as JPEG) and video” [Link] . Analytically, the DCT of an
M × N image can be defined via a cosine transform along each dimension (we omit
the formula here for brevity).

The Discrete Sine Transform (DST) similarly uses sine basis functions and is also a
Fourier-related transform. Like the DCT, a DST represents an image as a weighted
sum of sinusoids, but with odd symmetry. In practice, DSTs are less common than
DCTs, but they arise in certain boundary-processing contexts. In essence, both DCT
and DST are variants of the DFT that operate on purely real data with symmetric
(DCT) or antisymmetric (DST) extension of the image edges [Link] [Link] .

4. Image Enhancement

Image enhancement methods aim to improve visual quality or emphasize features.

Techniques are categorized by domain (spatial or frequency) and by objective
(contrast or detail).

4.1 Spatial-Domain Methods

4.1.1 Contrast Enhancement

Linear contrast stretching: Expand the range of intensities. If f (x, y) originally

lies in [Imin , Imax ], a linear stretch maps
​ ​

(f (x, y) − Imin )(L − 1) ​

g(x, y) = ​

Imax − Imin
​ ​

to utilize the full range [0, L − 1]. This increases global contrast if the image
uses only a narrow portion of the available gray levels.

Nonlinear transforms: Logarithmic or power-law (gamma) mappings can

enhance detail in dark or bright regions. For example, g = c log(1 + f )

compresses high intensities (highlighting details in shadows), and g = f

γ
(with
γ < 1 ) can brighten dark regions while compressing highlights. These
adjustments are often done pointwise with carefully chosen parameters.

Histogram equalization: A powerful automatic method uses the image’s

histogram [Link] . Let pf (i) be the normalized histogram (probability) of
​

i
intensities i. Define the cumulative distribution T (i) = ∑
k=0
​
pf (k)
​
. The
equalized output is

g(x, y) = round [(L − 1) T (f (x, y))].

This mapping flattens the histogram so intensities span the full range. As
Wikipedia notes, histogram equalization increases the global contrast especially
when the original image has intensities concentrated in a narrow range
[Link] . The result tends to brighten mid-tones and reveal detail in under- or
over-exposed regions. A drawback is it may boost noise, but it is very effective
for scientific images (e.g. X-rays) [Link] [Link] .

4.1.2 Local Histogram Methods

Instead of a single global histogram, one can apply equalization or contrast

adjustment locally (small regions) to adaptively enhance contrast (adaptive
histogram equalization). This helps when lighting is uneven. Simple thresholding of
histograms (e.g. minimum or mean threshold) is another basic segmentation-related
operation (see Section 6).

4.1.3 Spatial Smoothing (Low-pass Filtering)

Smoothing filters reduce noise and small details by averaging neighboring pixels.
Common methods include:

Image averaging: If multiple noisy frames of the same scene are available,
averaging them pixel-wise reduces random noise (by N ​
factor).

Mean filter (box filter): Replace each pixel by the average of itself and its
neighbors in a window (e.g. 3×3 or 5×5). This is a simple linear low-pass filter
that blurs the image [Link] .

Gaussian filter: A weighted average with a Gaussian kernel suppresses high-

frequency noise smoothly. It preserves edges better than a uniform mean filter.

In practice, these filters are implemented by convolution: g(x, y) =

∑
(s,t)∈W
​
f (x + s, y + t)h(s, t) where h is the mask (e.g. all 1’s for mean,
Gaussian weights, etc.). As GeeksforGeeks notes: “Low pass filter smoothens the
image by averaging the pixel of neighbouring value, thus minimizing random noise”
[Link] . Such filters achieve spatial domain smoothing directly.

4.1.4 Spatial Sharpening (High-pass Filtering)

Sharpening enhances edges and fine detail by accentuating high-frequency content.

Techniques include:

High-pass filters: Convolve with a mask that has a strong positive center and
negative neighbors (e.g. Laplacian). This highlights areas of rapid intensity
change (edges). For instance, the discrete Laplacian operator

2
∇ f = f (x + 1, y) + f (x − 1, y) + f (x, y + 1) + f (x, y − 1) − 4f (x, y)

yields positive values at edges (assuming dark to bright transition).

High-boost filtering: Amplifies the original image while subtracting a blurred

(low-pass) version. Formulaically: g(x, y) = f (x, y) + k [f (x, y) −

fLP (x, y)]

​
, where fLP is a smoothed version (low-pass) and k
​
> 0 . When k =

1 this reduces to the Laplacian. Larger k keeps more of the original detail. This
can be seen as adding back high-frequency content.

Derivative filters: First-order derivative masks (Sobel, Prewitt, Roberts) detect

edges by approximating gradients. Although often used for segmentation, they
can be used to sharpen by adding the gradient magnitude to the image (or using
unsharp masking).
In summary, sharpening filters boost high-frequency components. GeeksforGeeks
notes: “High pass filter is used to sharpen images; examples: Laplace filter and high
boost filter” [Link] . These filters are again implemented via convolution in the
spatial domain.

4.1.5 Homomorphic Filtering

Homomorphic filtering is a technique that works in the log-intensity domain to

normalize brightness and enhance contrast [Link] . The idea uses the
illumination-reflectance model f (x, y) = i(x, y) ⋅ r(x, y) . Taking the log makes it
additive: ln f = ln i + ln r . In frequency terms, illumination variations are mostly
low-frequency and reflectance (detail) is high-frequency. A homomorphic filter
suppresses the low-frequency (illumination) component and boosts high-frequency
(reflectance), then exponentiates back. As Wikipedia explains, it “simultaneously
normalizes the brightness across an image and increases contrast” by log-
transforming and high-pass filtering [Link] [Link] . Practically, one applies a
high-pass filter to the log of the image, then exponentiates. This is useful when
lighting is uneven: it darkens bright areas and brightens shadows. Homomorphic
filtering is effectively a nonlinear enhancement combining dynamic range
compression and contrast enhancement.

4.2 Frequency-Domain Methods

In the frequency domain, filtering is performed by multiplying the image’s DFT by a

filter function. One advantage is the ready implementation of ideal filters and the
convolution theorem. Common frequency filters include:

Low-pass filters (LPF): Ideal (rectangular), Butterworth or Gaussian LPF in the

frequency domain remove high frequencies, smoothing the image. E.g. an ideal
LPF of cutoff radius D0 sets F (u, v) to zero for
​
u
2
+ v
2
​
> D0 .
​

High-pass filters (HPF): Conversely, an ideal or Butterworth HPF suppresses low

frequencies, sharpening edges. For example, an ideal HPF sets F (u, v) = 0 if
u
2
+ v
2
​
< D0 . By the shift property, the DC component F (0, 0) (average
​

brightness) is removed, emphasizing variations.

The processing pipeline is: compute 2D DFT of the image, multiply by the chosen
frequency transfer function (LPF or HPF), then apply the inverse DFT to get the
filtered image. This is equivalent to convolution in the spatial domain [Link] .
GeeksforGeeks notes exactly this approach: “To simplify the process, we perform
multiplication in the frequency domain… transform to frequency domain (DFT), multiply
with filter, then go back via IDFT” [Link] . In practice, Gaussian and Butterworth
filters are favored because they avoid ringing artifacts (Gibbs phenomenon)
associated with ideal filters.

Thus, spatial and frequency methods are dual: averaging in space corresponds to
attenuating high-frequency components in the spectrum. Edge-detecting or high-
boost filters correspond to high-pass filters in frequency. Both domains provide a rich
toolbox: some enhancements (like homomorphic) combine both by log-transforming
(nonlinear) then filtering.

5. Image Restoration

Image restoration seeks to undo known or modeled degradations (blur, noise) in an

observed image. The degradation model is typically linear and shift-invariant:

g(x, y) = f (x, y) ∗ h(x, y) + n(x, y),

where f (x, y) is the original image, h(x, y) is the point spread function (PSF) of blur
(e.g. motion blur, defocus), and n(x, y) is additive noise. In the discrete Fourier
domain, this becomes

G(u, v) = F (u, v) H (u, v) + N (u, v).

Restoration methods attempt to estimate F from G given (or estimating) H and

statistical assumptions about n.

5.1 Algebraic Restoration (Inverse Filtering)

Unconstrained inverse filtering: If H (u, v) is known and non-zero, one can form
the inverse filter F (u, v) = G(u, v)/H (u, v) . Inverse filtering is optimal with
no noise, but in practice noise or small values of H (where ∣H ∣ ≈ 0 ) amplify
errors.
 Wiener filtering (minimum MSE): Incorporates a statistical model of noise vs.
signal. The Wiener filter in frequency domain is
∗
H (u, v)
^
F (u, v ) = ​

G(u, v),
2 Sn (u,v)
​

∣H (u, v)∣ + ​

Sf (u,v)
​

minimizing mean-square error. Sn , Sf are noise/image power spectra. This is an

​ ​

example of a constrained (regularized) solution.

Constrained least-squares: Here one minimizes ∥g − h ∗ f∥

2
+ λ∥Lf ∥
2
,
where L is a high-pass operator (e.g. Laplacian) enforcing smoothness. The
solution is a kind of regularized inverse (Tikhonov). These methods require
choosing a regularization parameter λ to balance fidelity vs. smoothness. (See
Gonzalez & Woods for derivations.)

5.2 Restoration by Homomorphic Filtering

Homomorphic filtering can also be viewed as a restoration approach when

illumination is treated as a degradation (multiplicative noise). By taking logs,
illumination gradients become additive noise that can be filtered. Essentially the
same process described in enhancement (log-transform, HPF, exponentiate) is used
to reduce multiplicative non-uniform illumination and thereby “restore” a more
uniformly lit image [Link] [Link] .

5.3 Geometric (Spatial) Restoration

Sometimes restoration involves correcting geometric distortions (e.g. lens distortion,

mis-registration). A geometric transformation re-maps pixel locations via a warping
function (affine, projective, lens correction, etc.). After determining the mapping from
distorted coordinates to true coordinates, one must interpolate pixel values onto a
regular grid.

Gray-level interpolation methods include nearest-neighbor, bilinear, or bicubic

interpolation. As one source notes, a geometric transform consists of (1) a spatial
transformation (rearranging pixel positions) and (2) gray-level interpolation to assign
intensities in the new image [Link] . For example, if an image must be rotated by a
non-integer angle, the resulting pixel grid does not align with original pixels;
interpolation (often bilinear for smoothness) is used to estimate the new pixel values.
These steps restore the image geometry, such as unskewing or aligning images.
6. Image Segmentation

Image segmentation partitions an image into meaningful regions (objects,

background) by detecting boundaries or region homogeneity. Key techniques include
point/edge detection, thresholding, and region-based methods.

6.1 Point, Line, and Edge Detection

Point detection: Identifies isolated points or noise (outliers). For example, the
Laplacian of an image will be large only at isolated bright or dark pixels. A simple
3 × 3 mask that responds strongly to a single bright pixel can act as a point
detector.

Line detection: Uses oriented masks (e.g. horizontal, vertical, 45°) to detect line
segments of different orientations. These are small convolution kernels that
respond to straight lines (length ≥ 3) in specific directions.

Edge detection: Finds boundaries between regions by locating large intensity

gradients. Common operators (Sobel, Prewitt, Roberts) compute approximations
of the image gradient ∇f . For instance, the Sobel operator uses two 3 × 3
∂f ∂f
masks to compute ∂x
​

and ∂y
​

; the gradient magnitude 2

fx + fy
​
2
​

highlights

edges. The Laplacian of Gaussian (LoG) or Difference of Gaussian (DoG) can

detect edges by zero-crossings of second derivatives. In practice, one often
smooths (with Gaussian) then takes derivatives to reduce noise.

Edge detectors are typically applied to the image and then thresholded to produce a
binary edge map. Combined detection (using multiple filters) can detect joined
features (e.g. corner detectors like Harris combine gradients).

6.2 Edge Linking and Boundary Detection

After edge points are detected, one may link them into continuous boundaries. Local
linking uses neighborhood criteria (connected edge pixels). Global methods seek
parameterized curves or shapes.

One powerful global method is the Hough Transform, which detects shapes (lines,
circles) by mapping points into a parameter space. For straight lines, each edge point
(x, y) votes in (ρ, θ)-space via the line equation ρ = x cos θ + y sin θ . Peaks in
this accumulator indicate dominant lines, effectively linking colinear edge points
[Link] . (For circles, one uses a 3D parameter space (a, b, r).) The Hough
transform thus performs boundary detection by solving a global voting problem
rather than local neighbor linking.

6.3 Thresholding

Thresholding separates an image into background and object by a gray-level cutoff T .

The simplest global thresholding produces a binary image g(x, y):

1, f (x, y) > T ,
g(x, y) = { ​ ​

0, f (x, y) ≤ T ,

for chosen T . (Or vice versa, depending on convention [Link] .) This segments
the image into two classes. In practice, one often normalizes intensities first.
Adaptive thresholding allows T to vary over the image (local windows) to account for
uneven lighting.

Optimal thresholding (Otsu’s method): An automatic global threshold can be found

by maximizing inter-class variance. Otsu’s algorithm tests all possible T and chooses
the one that maximizes the between-class variance (or equivalently minimizes intra-
class variance) of the two groups [Link] . This ensures the chosen T best
separates the histogram modes of foreground/background. As a result, Otsu’s
method “maximizes the between-class variance” to find the “optimum” threshold
[Link] .

6.4 Region-Oriented Segmentation

Region methods segment by grouping pixels with similar attributes:

 Region growing: Start from one or more seed pixels and add neighboring pixels
that have similar intensities (or other criteria) until no more can be added.
Similarity might be based on intensity difference below a threshold or statistical
homogeneity (variance). This grows regions that are locally uniform.

Split-and-merge: The image is recursively split (e.g. into quadrants) until each
region meets a homogeneity criterion (e.g. intensity variance below a threshold).
Then adjacent similar regions are merged. Wikipedia describes this as splitting
into quadrants and then merging adjacent homogeneous quadrants [Link] .
The result is a segmentation where each final region satisfies the homogeneity
condition.

These methods ensure that segmented regions are internally consistent. Split-and-
merge explicitly uses a quadtree structure and a predefined test (mean or variance) to
decide when to stop splitting [Link] . Region growing is simpler for smooth
images but can suffer from seed selection sensitivity.

In summary, segmentation can be performed by edge-based (gradient) methods,

thresholding, or region-based strategies, often in combination. Edge methods detect
boundaries; thresholding provides a quick binarization; region methods ensure area
homogeneity. A practical system might use a mixture: e.g. threshold first, then refine
by region growing and edge linking.

Key formulas used: In segmentation, the binary thresholding rule is a basic formula
[Link] . Histogram equalization mapping T (i) uses the cumulative histogram
to spread intensities (see Section 4.1.3) [Link] [Link] . Geometric
transforms (Section 2.2) are given by linear matrix equations. The 2D DFT
(Section 3.4) is the double sum above [Link] . These formulas form the core
mathematical tools for DIP.

Sources: The above notes draw on standard DIP texts and references. Fundamental
concepts and definitions are cited from authoritative sources [Link]

[Link] [Link] [Link] [Link] [Link] [Link]

[Link] [Link] [Link] , ensuring accuracy for academic study.

Citations

Fundamental Steps in Digital Image Processing | GeeksforGeeks

[Link]
 Digital image processing - Wikipedia
[Link]

Digital image processing - Wikipedia

[Link]

Digital image processing - Wikipedia

[Link]

Vector quantization procedure:

[Link]
[Link]

Fundamental Steps in Digital Image Processing | GeeksforGeeks

[Link]

Fundamental Steps in Digital Image Processing | GeeksforGeeks

[Link]

Fundamental Steps in Digital Image Processing | GeeksforGeeks

[Link]

Computer Vision and Image Processing - Fundamentals and Applications | PDF

[Link]

Computer Vision and Image Processing - Fundamentals and Applications | PDF

[Link]

A Simple Image Model | PDF | Sampling (Signal Processing) | Digital Image

[Link]

A Simple Image Model | PDF | Sampling (Signal Processing) | Digital Image

[Link]

A Simple Image Model | PDF | Sampling (Signal Processing) | Digital Image

[Link]

A Simple Image Model | PDF | Sampling (Signal Processing) | Digital Image

[Link]

A Simple Image Model | PDF | Sampling (Signal Processing) | Digital Image

[Link]

A Simple Image Model | PDF | Sampling (Signal Processing) | Digital Image

[Link]
 A Simple Image Model | PDF | Sampling (Signal Processing) | Digital Image
[Link]

2D Transformations in Computer Graphics

[Link]

2D Transformations in Computer Graphics

[Link]

[Link]
[Link]
[Link]

Difference Between Uniform and Non-Uniform Quantization | GeeksforGeeks

[Link]

Difference Between Uniform and Non-Uniform Quantization | GeeksforGeeks

[Link]

(PDF) Digital image processing Pixel Neighbors

[Link]

(PDF) Digital image processing Pixel Neighbors

[Link]

(PDF) Digital image processing Pixel Neighbors

[Link]

Fourier transform — Basics of Image Processing

[Link]

The Discrete Fourier Transform

[Link]

Fourier transform — Basics of Image Processing

[Link]

Fourier transform — Basics of Image Processing

[Link]

Fourier transform — Basics of Image Processing

[Link]

Fourier transform — Basics of Image Processing

[Link]

Fourier transform — Basics of Image Processing

[Link]

Discrete cosine transform - Wikipedia

[Link]

Discrete sine transform - Wikipedia

[Link]

Histogram equalization - Wikipedia

[Link]

Histogram equalization - Wikipedia

[Link]

Histogram equalization - Wikipedia

[Link]

Fundamental Steps in Digital Image Processing | GeeksforGeeks

[Link]

Fundamental Steps in Digital Image Processing | GeeksforGeeks

[Link]

Homomorphic filtering - Wikipedia

[Link]

Homomorphic filtering - Wikipedia

[Link]

No Slide Title
[Link]

Optimum Global Thresholding Using Otsu’s Method | GeeksforGeeks

[Link]

Optimum Global Thresholding Using Otsu’s Method | GeeksforGeeks

[Link]

Split and merge segmentation - Wikipedia

[Link]

A Simple Image Model | PDF | Sampling (Signal Processing) | Digital Image

[Link]

All Sources

geeksforgeeks [Link] [Link] scribd tutorialspoint

[Link] researchgate [Link] [Link] [Link]