UNIT 4

1. Explain the steps in image registration (8 marks)

2. Write the Matching using scene coherence algorithm to find the correspondence
between the points of sensed and reference images. (8 marks)
3. Explain the algorithm to find the correspondence between the lines of sensed and
reference images. (8 marks)
4. Explain the region by shape matching to find the correspondence between sensed and
reference image. (8 marks)
5. Explain the performance evaluation of objective evaluation measures with not
requiring a reference image. (8marks)
6. Explain the performance evaluation of objective evaluation measures with reference
image. (8marks)
7. Explain Multiresolution based fusion using discrete wavelet transform and curvelet
transform. (16 marks)
8. Explain Region based fusion. (8marks)
9. Explain hybrid image registration of image fusion. (8 marks)

1. Explain the Steps in Image Registration (8 Marks)

Image Registration

Image registration is a fundamental image processing operation used to align two or more
images of the same scene taken at different times, sensors or viewpoints. The steps involved
in image registration are:

1. Image Pre-processing

Before registration, images undergo preprocessing to improve the quality and to make feature
extraction easier.
Typical preprocessing tasks include:

 Noise removal using filters (Gaussian, median, Wiener)

 Histogram equalization for contrast enhancement
 Smoothing or sharpening
 Intensity normalization
This step ensures both reference and sensed images are in comparable form

2. Feature Detection / Extraction

Distinctive and stable features are detected from both images.

Common image processing–based features include:

 Edges (using Sobel, Canny)

 Corners / interest points (Harris, SIFT, SURF, ORB)
 Regions (blob detection using LoG, DoG)
These features represent meaningful structures that will be used for matching.
3. Feature Matching / Correspondence Identification

In this step, features extracted from the sensed image are matched to corresponding features
in the reference image.
Matching uses:

 Feature descriptors (SIFT, SURF, BRIEF, ORB)

 Similarity measures:
o Sum of Squared Differences (SSD)
o Normalized Cross-Correlation (NCC)
o Euclidean distance between descriptors
The result is a set of corresponding point pairs between the two images

4. Selection of Transformation Model

Based on the matched feature points, a geometric transformation model is selected to map the
sensed image onto the reference image.
Common transformation models in image processing include:

 Translation
 Rotation
 Scaling
 Affine transformation
 Projective transformation (Homography)
 Non-rigid (B-spline, Thin Plate Spline)
Robust estimation methods like RANSAC are used to eliminate outliers.

5. Parameter Estimation

The parameters of the chosen transformation model are computed using the matched feature
pairs.
For example:

 Affine: rotation angle, scaling factor, shear, translation

 Projective: 8 homography parameters
Least squares estimation is commonly used to solve for transformation parameters.

6. Resampling and Interpolation

After applying the transformation, the image must be resampled to assign intensity values to
the new pixel positions.
Common interpolation techniques include:

 Nearest Neighbor (fast, but blocky results)

 Bilinear interpolation (smooth results)
 Bicubic interpolation (high-quality resampling)
This step produces the geometrically corrected image.
7. Registered Image Generation and Evaluation

Finally, the registered image is generated. The accuracy of registration is evaluated using:

 Mean Squared Error (MSE)

 Mutual Information (especially in multi-sensor images)
 Correlation Coefficient
If the error is high, feature matching or transformation estimation may be refined.

[Link] the Matching using Scene Coherence Algorithm to find the correspondence
between the points of sensed and reference images.

Matching Using Scene Coherence Algorithm (8 Marks)

Scene Coherence Algorithm is used in image registration to determine the correct

correspondence between feature points of the sensed image and the reference image by
exploiting the geometric coherence of the scene. It assumes that neighboring points in one
image will have corresponding neighbors in the other image, thus preserving spatial
relationships.

1. Extraction of Feature Points

 Detect corner points, interest points, or edges from the reference image and sensed
image.
 Typical detectors: Harris, SUSAN, SIFT, etc.
 Let the feature sets be:
o Reference image points → R={r1 ,r2,…,rn}
o Sensed image points → S={s1 ,s2 ,…,sm}

2. Initial Similarity / Distance Computation

 For each point in the reference image, compute a similarity measure with all points in
the sensed image.
 Common measures:
o Euclidean distance between descriptors
o Intensity pattern similarity
o Gradient-based descriptors
 This forms an initial correspondence list, but it may contain many mismatches.

3. Neighborhood Relationship Analysis

 For every feature point in the reference image, examine its local neighborhood
(spatially close points).
 For each potential match in the sensed image, compare its neighborhood structure.
 Coherence is checked based on:
o Relative distances
o Angles between neighboring points
o Ordering of neighbors
 This helps eliminate matches that do not preserve geometry.
4. Coherence Constraint Verification

 The algorithm enforces that geometric relationships must remain consistent across
the images.
 A match (ri,sj) is valid only if:

d (ri , rk) ≈ d (sj , sl) : for neighboring points rk and sl.

 If the structure of neighbors does not match, the candidate correspondence is

discarded.

5. Consistency Voting

 Each pair of candidate matches “votes” for other matches that preserve the same
geometric transformation.
 Pairs that receive many supporting votes are considered coherent.
 Incoherent matches are removed.

6. Transformation Estimation from Coherent Matches

 Using the set of coherent point pairs, a transformation model (affine, rigid, projective)
is estimated.
 Methods like least squares or RANSAC refine the transformation and remove
remaining outliers.

7. Final Correspondence Set

 Only those point pairs that satisfy:

o geometric consistency
o neighborhood coherence
o transformation support
are kept as the final correspondence between sensed and reference images.

8. Generation of Robust Matching

 The final correspondence list is used to align the sensed image with the reference
image.
 The scene coherence algorithm ensures robust, spatially consistent feature matching
even in noisy or distorted conditions.
[Link] the algorithm to find the correspondence between the lines of sensed and
reference images

Algorithm to Find Correspondence Between Lines of Sensed and Reference Images (8

Marks)

Finding line correspondences is an important task in image registration, especially when

images contain prominent linear structures such as roads, edges of buildings, or boundaries.
The algorithm matches lines in the sensed image to those in the reference image by
comparing geometric and structural properties.

1. Line Detection

 Detect strong and continuous lines in both images.

 Common methods include:
o Hough Transform
o Canny edge detection + line fitting
o LSD (Line Segment Detector)
 Output:
o Reference line set → LR = {R1 ,R2,…,Rn}
o Sensed line set → LS = {S1,S2,…,Sm}

2. Extraction of Line Attributes

For each detected line, compute important attributes such as:

 Line orientation (angle θ)

 Line length
 Midpoint and endpoints
 Line normal vector
 Distance of the line from image origin
These attributes form the basis of matching.

3. Formation of Line Descriptors

 Create a descriptor for each line based on

o angle,
o length ratio,
o orientation difference,
o positional relationship.
 This descriptor helps compare lines independent of translation and small noise.

4. Initial Matching Based on Similarity

 For every reference image line Ri, compute similarity with each sensed image line Sj.
 A typical similarity measure:

Sim (Ri,Sj) = w1∣θRi−θSj∣ + w2∣lenRi−lenSj∣ + w3d(midpoints)

Pairs with minimum similarity score are chosen as initial matches.
5. Geometric Consistency Check

 Line matching must preserve the geometric configuration of the scene.

 For every pair of matched lines, evaluate:
o Distance between parallel lines
o Angle between intersecting lines
o Relative position of lines
 If ∣Angle (Ri,Rk) – Angle (Sj,Sl)| is large, the match is rejected.

6. Grouping Into Coherent Sets

 Lines that maintain similar structure across images are grouped.

 Example: lines forming the same corner or rectangle must match as a group.
 Inconsistent or isolated line matches are removed.

7. Transformation Estimation Using Line Correspondence

 Using the coherent line matches, estimate the geometric transformation:

o Rigid (rotation + translation)
o Affine transformation
o Projective transformation
 Least squares or RANSAC improves robustness and removes wrong matches.

8. Final Correspondence Set Extraction

 After transformation estimation, verify which line matches truly fit the model.
 Lines that agree with the final transformation are accepted as correct correspondences.
 Output: a robust mapping between sensed and reference image lines.

[Link] the Region-by-Shape Matching to Find the Correspondence Between Sensed

and Reference Images (8 Marks)

Region-by-shape matching is a region-based image registration technique in which regions

extracted from the sensed image are matched with their corresponding regions in the
reference image based on shape properties. It is especially useful when the images contain
objects with distinctive geometric shapes.
1. Region Extraction

 Segment both sensed and reference images into meaningful regions.

 Segmentation methods used:
o Thresholding
o Region growing
o Watershed segmentation
o Clustering-based segmentation (K-means, FCM)
 Output:
o Regions in reference image: R = {R1, R2,…}
o Regions in sensed image: S = {S1, S2,…}

2. Shape Representation

For each region, extract a consistent representation that captures its geometric structure.
Common shape representations:

 Boundary-based:
o Chain code
o Curvature signature
o Fourier descriptors
 Region-based:
o Moments (geometric moments, Hu moments)
o Area, centroid
o Compactness, eccentricity
These descriptors must be invariant to translation, rotation, and scaling.

3. Shape Descriptor Extraction

Compute shape descriptors for each region. For example:

 Area: number of pixels in the region

 Perimeter: length of region boundary
 Hu moment invariants: for scale/rotation invariance
 Aspect ratio: width to height
 Circularity: 4π⋅Area / Perimeter2

These descriptors are used for matching regions across images.

4. Similarity Measurement

Measure the similarity between each sensed-image region Sj and reference-image region Ri.
Common similarity measures:

 Euclidean distance between shape descriptor vectors

 Hausdorff distance between region boundaries
 Correlation between moment invariants

Lower distance → higher shape similarity.

5. Establishing Region Correspondence

For each region in the reference image, select the region in the sensed image that has the
highest similarity (minimum distance).
A candidate match must satisfy:

D (Shape (Ri) , Shape (Sj) ) < T ; where T = similarity threshold.

6. Spatial Consistency Verification

Since shapes must appear in similar spatial configurations, check:

 Relative position between neighboring regions

 Orientation of neighboring shapes
 Connectivity or adjacency
Regions that break the spatial structure are removed as false matches.

7. Transformation Estimation

Using the matched region centroids, estimate the global transformation:

 Translation
 Rotation
 Scaling
 Affine transformation
Centroids of matched shapes are input to least squares or RANSAC to compute
transformation parameters.

8. Final Region Correspondence

After transformation estimation, re-evaluate region matches.

If a region does not align well after mapping, it is discarded.
The final set of region correspondences is robust and geometrically consistent across both

images.
[Link] the performance evaluation of objective evaluation measures without
requiring a reference image. (8 marks)

Performance Evaluation of No-Reference Objective Quality Measures

No-reference (NR) image quality assessment (IQA) evaluates the quality of an image
without using an original, undistorted reference image. These methods analyze the
distorted image itself and predict its quality based on statistical, structural, or perceptual
characteristics.

1. Natural Scene Statistics (NSS)–Based Measures

 Natural images follow predictable statistical distributions (mean, variance, gradients,

wavelet coefficients).
 Distortions such as blur, noise, or compression introduce deviations.
 NR-IQA methods evaluate quality by measuring how much the image deviates from
natural statistical behavior.
 Example measures:
o BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator)
o NIQE (Natural Image Quality Evaluator)
 Lower deviation from NSS ⇒ better quality.

2. Sharpness/Blur-Based Measures

 Many no-reference methods measure loss of high-frequency content.

 Techniques:
o Gradient magnitude analysis
o Edge width analysis
o Laplacian variance
 Less edge strength → higher blur → lower quality.
 Useful for evaluating out-of-focus images.

3. Noise Estimation Measures

 These methods detect random variations in pixel intensity.

 Techniques:
o Local variance estimation
o Wavelet-domain noise analysis
o High-frequency residual estimation
 Higher estimated noise indicates poorer image quality.

4. Blockiness and Compression Artifact Measures

 In JPEG/JPEG2000 compression, block boundaries become visible.

 NR-IQA computes:
o Difference between adjacent 8×8 blocks
o Regular block patterns
o Quantization noise features
 More blockiness or ringing → lower quality.
 Used widely in video/image compression evaluation.
5. Contrast and Intensity-Based Measures

 Poor quality images often suffer from:

o Low contrast
o Illumination variation
o Saturation/underexposure
 Metrics evaluate histogram characteristics:
o Dynamic range
o Entropy
o Local contrast
 Higher contrast and balanced intensity → better perceived quality.

6. Texture and Structure-Based Measures

 Distortion affects structural information (edges, lines, textures).

 NR measures analyze:
o Local binary patterns (LBP)
o GLCM-based texture features
o Structural degradation in repetitive patterns
 Loss of texture consistency implies low-quality images.

7. Machine Learning / Deep Learning Based NR-IQA

 Features extracted from the distorted image are fed into models that predict
perceived quality.
 Methods:
o Support Vector Regression (SVR)
o CNN-based quality predictors
 Trained on large IQA datasets to map distortions → quality scores.
 Provide high accuracy compared to traditional NR methods.

8. Perceptual Quality Estimation (Human Visual System Based)

 Models simulate the Human Visual System (HVS).

 Evaluate:
o Visual masking
o Sensitivity to edges
o Saliency detection
o Contrast sensitivity
 Distortions in perceptually important regions reduce quality significantly.
 Produces a perceptually aligned quality score even without a reference.

Conclusion

Objective evaluation measures without reference images rely on analyzing the internal
characteristics of the distorted image—statistics, sharpness, noise, texture, and perceptual
cues—to estimate its quality. They are highly useful in real-time applications like
surveillance, enhancement, and compression where reference images are unavailable.
[Link] the performance evaluation of objective evaluation measures with reference
image. (8marks)

Performance Evaluation of Objective Measures With Reference Image (8 Marks)

Full-reference (FR) objective evaluation measures compare the distorted (test) image with
the original reference image to quantify the loss of quality. These measures assume that the
reference image is perfect and provide numerical scores indicating how close the distorted
image is to the original.

1. Mean Squared Error (MSE)

 MSE measures the average squared difference between reference and distorted image
pixels.

MSE =1/ MN ∑ i=1 M ∑j=1 N (R(i,j) − D(i,j)) 2

Lower MSE → less distortion → higher quality.

 Simple but does not correlate strongly with human perception.

2. Peak Signal-to-Noise Ratio (PSNR)

 PSNR is derived from MSE and widely used for evaluating compression quality.

PSNR = 10log10 (L2/MSE)

where L is maximum pixel value (e.g., 255).

 Higher PSNR indicates better fidelity.

 Good for measuring overall pixel-level differences.

3. Structural Similarity Index (SSIM)

 SSIM evaluates visual structure, not pixel differences.

 Uses three components:
o Luminance similarity
o Contrast similarity
o Structural similarity

SSIM(x,y) = [ l(x,y)]α [c(x,y)]β [s(x,y)]γ

SSIM = 1 → perfect match.

 Strong correlation with human visual perception.

4. Multi-Scale SSIM (MS-SSIM)

 Evaluates image similarity across multiple resolutions.

 Better at detecting: Blur, Noise, Compression artifacts
  Closer to real human visual system performance than SSIM.

5. Feature Similarity Index (FSIM)

 Measures similarity using phase congruency and gradient magnitude, which are
important visual features.
 Sensitive to structural distortions and blurring.
 High FSIM value → high visual similarity.

6. Visual Information Fidelity (VIF)

 Based on natural scene statistics and information theory.

 Estimates how much information from the reference image is preserved in the
distorted image.
 Higher VIF → more information retained → better quality.

7. Universal Image Quality Index (UIQI)

 Measures similarity in terms of:

o Correlation
o Contrast
o Luminance
 UIQI ranges from −1 to 1.
 Higher values indicate better quality.

8. Perceptual Evaluation (HVS-Based Measures)

 Uses models of the Human Visual System (HVS) to evaluate:

o Contrast sensitivity
o Visibility of distortions
o Importance of edges
 More accurate for perceptual image quality scoring.
 Examples:
o HDR-VDP (High Dynamic Range Visual Difference Predictor)
o HVS-based contrast models

Conclusion

Full-reference objective evaluation measures compare the test image with a perfect reference
image and compute how much distortion has occurred. Measures like PSNR and MSE give
mathematical accuracy, while SSIM, FSIM, VIF, and HVS-based models correlate better
with human visual perception.

[Link] Multiresolution based fusion using discrete wavelet transform and curvelet
transform. (16 marks)

A. Introduction to Multiresolution-Based Image Fusion

Image fusion combines information from multiple images (e.g., multispectral +

panchromatic, medical modalities, sensor pairs) to produce a single image with higher quality
and more useful information.
Multiresolution fusion uses transform-domain methods (Wavelet, Curvelet) to represent
image features at different scales and orientations before fusing them.

B. Discrete Wavelet Transform (DWT) Based Fusion (8 marks)

Discrete Wavelet Transform (DWT) based image fusion decomposes input images into low-
frequency (approximation) and high-frequency (detail) subbands using DWT. The
corresponding subband coefficients are then combined using fusion rules such as maximum
selection (for details) and averaging (for approximation). Finally, the fused image is
obtained by applying Inverse DWT to the fused coefficients, producing an image with
enhanced clarity and preserved edges.

1. Concept of DWT

 DWT decomposes an image into one low-frequency (approximation) component

and three high-frequency (detail) components:
o LL → Approximation (low-low)
o LH → Horizontal details
o HL → Vertical details
o HH → Diagonal details
 This provides a multiresolution representation similar to how the Human Visual
System perceives images.

2. Steps in DWT-Based Fusion

Step 1: Apply DWT to source images

For input images AAA and BBB:

DWT(A) = {LLA, LHA, HLA, HHA}

DWT(B) = {LLB, LHB, HLB, HHB}
Step 2: Fusion Rules

Fusion rules may include:

 Maximum selection rule (for details)

Cdetail (i,j) = max (Adetail (i,j) , Bdetail (i,j))

Averaging rule (for approximation)

CLL = (LLA + LLB) / 2

 Energy-based rule
o Compare local energies and select the coefficient with larger energy.

Step 3: Construct fused wavelet coefficients

Combine fused LL, LH, HL, HH to form:

C = {CLL , CLH , CHL , CHH}

Step 4: Apply Inverse DWT
F = IDWT(C) , This gives the final fused image.

3. Advantages of DWT Fusion

 Preserves edges and textures.

 No spatial distortion due to perfect reconstruction filters.
 Good for compression, denoising, and medical image fusion.

4. Limitations of DWT

 Cannot capture directional features well.

 Only three orientations (H, V, D).
 Limited in handling curved edges and textures.

C. Curvelet Transform Based Fusion (8 marks)

Curvelet Transform is an advanced multiresolution transform designed to overcome DWT’s

limitations. Curvelet Transform based image fusion decomposes images into multiple scales
and directions using curvelets, which efficiently represent edges and curves. The
corresponding curvelet coefficients from the input images are fused using rules like
maximum coefficient selection, and the inverse Curvelet Transform reconstructs a fused
image with improved edge details and directional features.

1. Concept of Curvelet Transform

 Provides multiscale and multidirectional representation.

 Captures curved edges efficiently (wavelets fail here).
 Decomposition yields many directional subbands, unlike wavelets.

2. Steps in Curvelet-Based Fusion

Step 1: Apply Curvelet Transform

For input images AAA and BBB:

CT(A)={CAs,θ(x,y)}
CT(B)={CBs,θ(x,y)} ; where

 s → scale
 θ → direction
 Cs,θ → curvelet coefficients

Step 2: Fusion of Curvelet Coefficients

Use rules such as:

a) Maximum absolute value

b) Energy-based fusion

 Calculate local energy of coefficients

 Choose coefficient with higher energy

c) Weighted fusion

Weights based on local variance or entropy.

Step 3: Form Fused Curvelet Coefficients

Combine fused coefficients across all scales and directions.

Step 4: Inverse Curvelet Transform

F = CT−1(CF)

Produces final fused image with enhanced structural information.

3. Advantages of Curvelet Transform

 Better at capturing edges along curves (roads, vessels, boundaries).

 Better directional sensitivity (many orientations).
 Produces sharper fusion results than DWT.
 Excellent for medical, satellite, and multisensor fusion.

4. Limitations

 Computationally expensive.
 More complex to implement.
 Requires more memory compared to wavelets.

D. Comparison: DWT vs Curvelet Transform (for conclusion)

Feature DWT Curvelet

Edge Poor for

Excellent
representation curves

Directions Only 3 Many orientations

Computation Fast Slower

 Feature DWT Curvelet

Fusion quality Moderate High

Medical, satellite, high-detail

Applications Basic fusion
fusion

[Link] Region based fusion. (8marks)

Region-Based Fusion

Region-based fusion is an image fusion technique in which the input images are divided into
meaningful regions, and fusion is performed at the regional level instead of pixel or
coefficient level. This method preserves structural information better and provides improved
visual quality.

1. Segmentation of Input Images

 Both source images (e.g., IR, visible, CT, MRI) are segmented into regions.
 Segmentation can be done using:
o Region growing
o Watershed
o Clustering (K-means, Fuzzy C-means)
o Edge-based segmentation

The goal is to divide images into homogeneous and meaningful regions.

2. Region Labeling

 After segmentation, each region is assigned a unique label.

 Corresponding regions between two images are identified based on:
o Position
o Similarity
o Shape
 Helps in matching which region of Image A corresponds to Image B.

3. Extraction of Regional Features

For every region, features are extracted such as:

 Mean intensity
 Variance
 Texture features
 Edge strength
 Energy

These features decide which region contains more useful information.

4. Region-Level Fusion Rules

Fusion is done based on the extracted features. Common rules include:

a) Maximum energy rule

Region with higher energy/variance contributes more.

b) Salient-region selection

Choose the region that contains:

 More edges
 Higher contrast
 More texture details

c) Weighted averaging

Combine corresponding regions proportionally.

d) Decision maps

Binary or multi-level maps decide from which image each region should be taken.

5. Construction of the Fused Image

 After selecting or fusing regions from both images, all fused regions are combined.
 Region boundaries are smoothed to avoid artifacts.
 Final fused image contains:
o High-intensity regions
o Strong edges
o Preserved spatial information

6. Advantages

 Preserves important structures (edges, objects).

 Reduces noise by considering region statistics.
 More perceptually meaningful than pixel-level fusion.

7. Applications

 Medical image fusion (CT + MRI)

 Remote sensing
 Night vision enhancement
 Surveillance
Conclusion

Region-based fusion uses segmentation and region-level decision making to produce a fused
image with enhanced detail and structural clarity.

[Link] hybrid image registration of image fusion. (8 marks)

Hybrid Image Registration in Image Fusion

Hybrid image registration combines multiple registration techniques—such as feature-

based, intensity-based, and transformation-based methods—to align images more accurately
before performing image fusion. It takes advantage of the strengths of each method while
minimizing their weaknesses.

1. Introduction

Hybrid image registration is used when images come from different sensors, different
modalities (CT–MRI, visible–IR), or when the images have large variations in scale,
intensity, or geometry.
It improves alignment accuracy, which is essential for high-quality image fusion.

2. Combined Use of Feature-Based and Intensity-Based Methods

Feature-Based Component

 Extracts key features: edges, corners, lines, landmarks.

 Uses detectors like:
o Harris corner
o SIFT / SURF
o Hough lines
 Good when intensity changes across modalities (e.g., CT vs MRI).

Intensity-Based Component

 Uses pixel intensity similarity.

 Measures like:
o Mutual information (MI)
o Correlation coefficient
 Helps refine alignment after coarse feature matching.

3. Steps in Hybrid Registration

Step 1: Preprocessing

 Noise removal
 Contrast enhancement
 Normalization

Step 2: Feature Extraction

  Detect stable features (points or edges).
 Select region-of-interest if necessary.

Step 3: Initial Alignment (Coarse Registration)

 Performed using feature-based matching.

 Estimates transformation parameters like:
o Translation
o Rotation
o Scaling

Step 4: Fine Alignment (Intensity Refinement)

 Apply intensity-based optimization:

o Mutual Information
o Gradient descent
o Correlation optimization
 This step reduces small mismatches.

Step 5: Transformation Model

Hybrid methods may use:

 Affine transform
 Projective transform
 Spline-based (nonrigid) transform
 B-spline or thin-plate spline for deformable images

Step 6: Interpolation & Resampling

 Final registered image is resampled using:

o Bilinear
o Bicubic
o Sinc interpolation

Step 7: Fused Image Generation

 Once aligned, images can be fused using:

o DWT fusion
o PCA fusion
o Region-based fusion
o Curvelet fusion

4. Advantages of Hybrid Registration

 Robust against noise and geometric distortions.

 Accurate alignment for multimodal images.
 Better performance than using a single method.
 Works for rigid and non-rigid image pairs.
5. Applications

 Medical imaging (CT–MRI, PET–CT).

 Remote sensing images from different satellites.
 Surveillance and IR–visible fusion.
 Robotics and autonomous navigation.

Conclusion

Hybrid image registration integrates feature-based matching and intensity-based refinement

to achieve high-precision alignment. This improved registration accuracy directly enhances
the quality of fused images by ensuring that corresponding structures are properly aligned.

2 MARKS

[Link] are the steps to follow the image registration?

The main steps for image registration are feature detection and description, feature matching,
outlier rejection, transformation estimation, and image resampling. First, detect and describe
distinctive features in both images, then match these features to find correspondences. Next,
filter out incorrect matches (outlier rejection) to accurately estimate the geometric
transformation model, and finally, use this model to resample and align the images.

[Link] image registration.

Image registration is the process of aligning two or more images to a common coordinate
system, making their corresponding features match spatially. This is necessary to compare or
integrate data from different sources, such as images taken at different times, from different
viewpoints, or with different sensors, in order to facilitate analysis and combine their
information.
Purpose: It establishes a spatial correspondence between images, which is essential for tasks
like image fusion, tracking changes over time, and creating detailed 3D models from multiple
2D images.
Process: One image is designated as the reference, and a geometric transformation is applied
to the other(s) to align them with the reference.

[Link] inverse filter. What are the uses?

An inverse filter is a technique in image processing that restores a degraded image by

applying the mathematical inverse of the blurring function to it. Its main use is image
restoration, or deblurring, by undoing distortions like motion blur or out-of-focus effects, but
it is highly sensitive to noise, which can be significantly amplified. This makes it best for
scenarios where the blur is known and there is minimal noise, with more advanced filters like
the Wiener filter often used when noise is present.

Uses:
Image Restoration (Deblurring): Its primary use is to deblur an image that has been blurred
by motion, camera shake, or other optical effects.
 Image Enhancement: By reducing blur, it improves the visual quality of an image, making
details sharper and more visible.
Specific Applications:
Astronomy: To deblur images of celestial objects that are affected by atmospheric distortion.

Law Enforcement: To enhance images for forensic analysis.

Medical Imaging: To improve the clarity of medical scans for better diagnosis.

[Link] cornerness measure.

A cornerness measure is a mathematical function or algorithm in image processing that

assigns a numerical score to each pixel (or image patch) to quantify how likely that region is
to be a corner or a feature of interest.
The core idea is that a corner, when viewed within a small surrounding window, will exhibit
significant changes in intensity when that window is shifted in any direction (horizontal,
vertical, or diagonal). the goal of a cornerness measure is to assign a score that, when a
threshold is applied, allows the efficient and accurate identification of corners within an
image.

[Link] Hausdorff distance

The Hausdorff distance in image processing is a measure of how far apart two sets of points
are, often used to compare the similarity between two shapes, like a predicted segmentation
and a ground truth, or a model and an image. It is defined as the maximum of all the
minimum distances from a point in one set to the nearest point in the other set. This makes it
sensitive to outliers but also useful for tasks where even a single mismatch is significant.

 The process involves comparing two sets of points, A and B, which could represent the
boundaries of objects in images.
 For each point in set A, the shortest distance to any point in set B is found. The Hausdorff
distance is the largest of these minimum distances.
 This process is performed in both directions: finding the largest minimum distance from A to
B, and the largest minimum distance from B to A. The final Hausdorff distance is the larger
of these two values.

[Link] is the principle to relate the points in sensed and references images?

The principle to relate points in sensed and reference images is image registration, which
aligns two images into a common coordinate system. This process involves identifying
corresponding points or features, known as tie points or control points, in both images and
then using a geometric transformation to map the sensed image onto the reference image.

 Definition: The process of overlaying two or more images of the same scene taken at
different times, from different viewpoints, or by different sensors.
 Goal: To align the images to a common coordinate system.
 [Link] Projective transformation.
A projective transformation, also known as a homography or perspective transformation,
is a non-linear geometric transformation in image processing that maps the points of an image
from one plane to another. It simulates the effect of a change in viewpoint, such as when a
flat object is viewed from an angle.
Key Characteristics
 Preserves collinearity: Straight lines in the original image remain straight lines after the
transformation.
 Does not preserve parallelism: Parallel lines in the original image may converge to a
vanishing point (as in perspective distortion), unless the transformation is a special case
known as an affine transformation.
 Does not preserve angles or lengths: Angles and the scale of objects can change non-
uniformly across the image.

[Link] are the uses of Invariant moments?

Invariant moments are widely used in image processing for tasks that require object
identification and shape analysis, particularly when the objects may be subject to various
transformations. Their primary uses include:
 Object Recognition: They allow for the identification of shapes regardless of their position,
size, or orientation in an image.
 Image Retrieval: In large databases, invariant moments can serve as shape descriptors to
search for images that contain similar objects.
 Medical Imaging Analysis: They assist in characterizing the shape of anatomical structures
or anomalies, such as tumors or lesions, which can aid in diagnosis.
 Character Recognition: The moments help in identifying handwritten or printed characters,
as the character's shape remains the same even if its position or size varies slightly.
 Texture Analysis: In addition to shape, invariant moments can provide information related
to an image's texture features.
 Motion Analysis and Tracking: By analyzing the moments of an object across successive
frames, its movement can be tracked, which is useful in surveillance or autonomous
navigation.

[Link] relaxation labelling

Relaxation labeling in image processing is an iterative, parallel process used to assign a

unique and consistent label to pixels (or objects) in an image by utilizing local constraints
and contextual information. It improves the initial, possibly ambiguous, labeling by favoring
classifications that are compatible with those of neighboring features.
 Goal: The primary goal is to reduce ambiguity in initial classifications and achieve a globally
consistent interpretation of the image.
 [Link] shape matrices.

In image processing, a "shape matrix" can refer to two concepts: either a matrix that
represents the pixel values of an image to allow for manipulation, or a matrix used for
geometric transformations like rotation, scaling, and translation of shapes. Digital images are
represented as matrices of pixels, where each pixel's value is stored in a cell of the matrix,
enabling operations like filtering, blurring, and sharpening through the use of smaller
matrices called kernels.
Image representation
 Pixel-based matrices: A digital image is represented as a matrix where each element
corresponds to a pixel.
 Matrix dimensions: The size of the matrix (m x n) corresponds to the image's dimensions
(e.g., height x width).
 Color images: A color image is typically represented by three matrices (layers), one each for
red, green, and blue (RGB), combined to create the final image.

[Link] Image fusion

Image fusion is a process in image processing that combines multiple input images into a
single, more informative composite image. This resulting image contains a more
comprehensive and enhanced representation of the scene than any of the original images
alone, which is useful for human or machine perception and further analysis. It can be
performed by combining spatial information, frequency domain information, or even
decisions made from individual images.
 Goal: To create a single image with higher information content by merging features from
multiple source images.
 Information integration: It combines complimentary data from different sensors,
viewpoints, or modalities.

[Link] multisensory fusion

Multisensory fusion in image processing is the process of combining information from

multiple image sources or sensors to create a single, enhanced composite image that is
more informative, accurate, and reliable than any of the individual source images. The goal is
to leverage the complementary information from different sensors to improve the overall
understanding of a scene or object for human or machine perception.
 Multiple Sources: The input involves images from different sensors or modalities (e.g., a
visible light camera, an infrared camera, a LiDAR sensor, MRI, CT scans).
 Enhanced Information: The resulting fused image integrates the most useful details from
each source. For example, fusing a high-spatial-resolution panchromatic image with a lower-
resolution multispectral image yields a single high-resolution, full-color image (known as
pansharpening).

[Link] are the potential advantages of image fusion?

 Image fusion in image processing offers several advantages by combining complementary
information from multiple source images into a single composite image [1]. This integration
leads to a more comprehensive and robust representation of the scene, with key benefits
including:
 Improved Clarity and Detail: Fusion enhances the overall clarity and amount of detail
present in the image.
 Enhanced Information Content: The resulting fused image typically contains more
complete and accurate information than any of the individual source images alone.
 Increased Reliability: By merging redundant information from multiple sources, the fused
image becomes more reliable and less susceptible to noise or artifacts present in just one of
the input images.
 Better Human and Machine Perception: The improved quality and information density
make images easier for both humans to interpret (e.g., for medical diagnosis or surveillance
analysis) and for machine vision systems to process (e.g., in feature extraction and scene
understanding).
 Improved Decision Making: With more comprehensive and reliable data, image fusion
facilitates better-informed decisions in various applications, such as medical diagnostics,
military target identification, and environmental monitoring.

[Link] are the applications of image fusion?

Applications of image fusion

 Medical Imaging: Fuse images from different modalities like MRI, CT, and PET to provide
a more complete view for diagnosis and surgical guidance.
 Remote Sensing: Combine images from different satellite sensors, such as a high-resolution
panchromatic image with a low-resolution multispectral image, to create a single image with
both high spatial detail and rich spectral information.
 Surveillance and Security: Integrate data from various sensors, such as visible and infrared
cameras, to create a more comprehensive picture for enhanced object identification and
tracking.
 Robot Vision: Combine information from multiple sensors, such as stereo cameras, to help
robots perceive their environment more effectively for navigation and manipulation tasks.
 Manufacturing and Diagnostics: Monitor processes and inspect products by fusing images
to improve the ability to identify defects or variations that may not be visible in a single
image.
 Photography: Create advanced photographic effects by fusing multiple images with different
exposures (exposure fusion) or different focal planes (multifocus fusion).
 Astronomy: Combine images of celestial objects from different telescopes or at different
wavelengths to create more detailed and informative composite images.

[Link] are the methods of image fusion?

 Spatial domain methods
 IHS (Intensity-Hue-Saturation) Transform: Converts an image to IHS space to add high-
resolution spatial details to the color information.
 Principal Component Analysis (PCA): A statistical method that transforms pixel values to
reduce redundancy while preserving image quality.
 High-pass filtering: Injects high-frequency details from one image into another.
 Averaging: A simple method that combines images by averaging their pixel values.
 Brovey Method: A technique often used for pan-sharpening, which fuses a multispectral
image with a panchromatic image.
Transform domain methods
 Pyramid techniques: Decompose images into different levels of resolution, then combine
the representations to create a composite image.
 Wavelet Transform: Decomposes images into approximation and detail coefficients, which
can then be fused based on specific rules before reconstruction.
 Discrete Cosine Transform (DCT): Another frequency-domain transform used for image
fusion.
 Curvelet Transform: A multi-resolution analysis tool that offers better performance than
wavelet methods in some cases.
Feature and decision level methods
 Feature level fusion: Extracts meaningful features such as edges or textures from input
images and combines them into a new set of features.
 Decision level fusion: Processes each image individually to make a decision or extract
information, and then fuses these decisions to get a final output.
 Deep learning methods: Uses deep neural networks to process features, fuse them, and then
reconstruct the final image.

[Link] are multiscale decomposition methods?

Multiscale decomposition in image processing refers to techniques that break down an image
into components at different resolutions, or "scales," to analyze and process its features. This
allows for separating coarse contours from fine details and is used for tasks like image fusion,
compression, denoising, and segmentation. Common methods include using Gaussian
pyramids, Laplacian pyramids, and wavelet transforms.
Common techniques
 Pyramid methods: These techniques create a hierarchy of images, where each subsequent
image is a lower-resolution version of the previous one, like Gaussian and Laplacian
pyramids.
 Wavelet transforms: This method decomposes an image into different frequency and spatial
components, providing a more detailed representation than pyramids alone.
 Empirical Mode Decomposition (EMD): This is a method used for analyzing non-
stationary and non-linear signals, which can be applied to image data.
 Patch-based decomposition: Some methods decompose images into patches and analyze
these patches at different scales for specific tasks like anomaly detection.

[Link] are non multiscale decomposition methods?

Non-multiscale decomposition methods in image processing are techniques that operate

directly on the original image at a single resolution or do not inherently break down the
image into a hierarchy of different scales or frequency bands. These methods often work in
the spatial domain or use certain transforms that do not rely on the creation of a scale-space
representation.
Common examples of non-multiscale decomposition methods include:
 Pixel-based methods: These techniques process individual pixels, often using simple
operations or comparisons, to produce a result.
o Minimum/Maximum techniques: In image fusion, these methods select the lowest or
highest intensity value for each corresponding pixel from the source images.

o Weighted Averaging: A method where each pixel in the source images is assigned a weight,
and the resulting image is produced by a weighted sum of the pixel values.

 Region/Block-based methods: These methods operate on defined blocks or regions of the

image.
o Simple Block Replace Technique: Involves adding pixel values within a block and taking
the average for the resultant block.

o Non-local block methods: Techniques that compare and process blocks of pixels that are
distant from each other but share similarities.

 Feature Space Transform-based methods:

o Hue Intensity Saturation (HIS) transform: A color image technique that converts an RGB
image into HIS components for processing (e.g., intensity level modification).

[Link] overall cross entropy.

In image processing, the overall cross-entropy typically refers to the average cross-entropy
loss calculated across an entire dataset during the training of a classification model, such as
a neural network. This metric quantifies the total difference between the predicted probability
distributions for all pixels or images and their corresponding true (ground truth) label
distributions.
Definition and Purpose
Cross-entropy (CE) is a measure from information theory that quantifies the difference or
"surprise" between two probability distributions, the true distribution (P) and the predicted
distribution (Q).
 Purpose: As a loss function, the goal of a model's training process is to minimize this value.
A lower cross-entropy value indicates that the model's predicted probabilities are closely
aligned with the actual labels, meaning better performance.

[Link] universal quality index.

The Universal Quality Image Index (UQI) is an image processing metric that models
distortion as a combination of three factors: loss of correlation, luminance distortion, and
contrast distortion. It is a value between 0 and 1, with 1 indicating high quality, and it
performs better than simpler metrics like mean squared error (MSE) across a variety of
distortion types, such as blurring, noise, and compression.
 Modeling distortion: The UQI is based on a model of image degradation that considers three
key factors:
o Loss of correlation: How similar the patterns in the two images are.

o Luminance distortion: The difference in overall brightness or intensity between the two
images.

o Contrast distortion: The difference in the range of pixel values (contrast) between the two
images.

 Value range: The index has a value from 0 to 1, where a value of 1 represents a perfect
match between the two images and lower values indicate greater distortion.
 Applications: It is used to evaluate image quality in various applications, including:
o Image compression

o Noise contamination

o Blurring

o Contrast stretching

[Link] mutual information

Mutual information (MI) in image processing is a statistical measure of the dependency

between the intensity values of corresponding pixels in two images. It quantifies how
much information one image provides about the other, or the average reduction in uncertainty
about one image's pixel values when the other image's pixel values are known.

Applications include:
 Multimodal image registration
 Image segmentation.
 Feature selection in machine learning applied to image analysis.
 Image fusion performance evaluation.
UNIT 5 – 2 MARKS
[Link] absorption rule
The "absorption rule" is not a standard, general rule in the field of digital image processing.
The term "absorption" in this context typically refers to the
physical process of light or other electromagnetic radiation transferring energy to
matter when forming an image or interacting with a sample.
This physical phenomenon is relevant in specific image processing applications, such as:

 Image Formation
 Medical Imaging
 Spectroscopy and Remote Sensing
 Image Enhancement
 Boolean Logic (Unrelated)

[Link] Voxel
A voxel is the 3D equivalent of a pixel, representing the smallest unit of a 3D image or
volumetric dataset. It is a cube-shaped element in a three-dimensional grid that contains a
value for a specific point in that volume, such as color or intensity, similar to how a 2D image
is made of pixels. Voxels are fundamental to applications like medical imaging (CT scans,
MRIs) and 3D computer graphics.
Key characteristics of a voxel:

 Volume element: A voxel represents a discrete volume of space, whereas a pixel

represents a 2D area.
 3D grid: Multiple voxels are arranged in a 3D grid to form a complete 3D image.
 Value: Each voxel is assigned a value based on the data at its location, such as an
intensity value in a medical scan.
 Image creation: In medical imaging, a series of 2D image slices are stacked together
to create a volumetric dataset where each pixel in a slice corresponds to a voxel.
 Applications: Voxels are used in medical imaging, 3D rendering, and computer
games to create and represent 3D objects and scenes.

[Link] are the difference between pixel and voxel

A pixel is the smallest unit in a 2D image, representing a single point with an X and Y
coordinate, while a voxel is the 3D equivalent, representing a point in a volume with X, Y,
and Z coordinates. Pixels make up 2D images (like photos on a screen), and voxels build 3D
images (like those from a medical scan), with a voxel's dimensions being a pixel's area plus
the third dimension, or slice thickness.
Feature Pixel Voxel

Dimensions 2D (X and Y axes) 3D (X, Y, and Z axes)

Definition A "picture element" A "volume element"

Representation A 2D square in a grid A 3D cube in a volumetric grid

 Examples Digital photos, computer Medical imaging (CT, MRI), 3D
monitor displays graphics

Relationship to other A pixel is the base for a A voxel is a pixel plus the third
unit voxel's area dimension (slice thickness)

[Link] to store 3D data sets

3D datasets in image processing are typically stored using a variety of formats and methods
depending on the data type (e.g., volumetric, mesh, point cloud) and application needs.
3D data sets in image processing are stored as 3D voxel arrays (volume data) or as a stack
of 2D image slices. They are commonly saved using formats like DICOM, RAW, NIfTI, or
TIFF stacks, which preserve the x–y–z spatial structure and intensity information of the 3D
volume.
Storage Formats:

o DICOM (Digital Imaging and Communications in Medicine): The widely used standard
for storing, transmitting, and managing medical images and related patient data.

o NIfTI (Neuroimaging Informatics Technology Initiative): A common format in

neuroscience for functional and structural brain imaging data.

o TIFF stacks: Multiple 2D TIFF images stacked to form a 3D volume.

 Point Clouds: A set of unstructured 3D coordinates (X, Y, Z) that represent the surface of an
object. Points may be augmented with additional information like color or surface normals.
o Storage Formats:

o PCD (Point Cloud Data): A file format specifically designed for storing point cloud data.

o PLY (Polygon File Format): A format that can store point clouds and associated data.

 Meshes
Key Software and Tools
 MATLAB
 3D Slicer
 Open3D
[Link] to obtain 3D data sets by different imaging techniques
3D datasets are acquired using various imaging techniques, which can be broadly categorized
into medical imaging modalities, surface scanning methods, and image-based reconstruction
techniques.
3D data sets are obtained by acquiring multiple 2D image slices from imaging techniques
such as CT, MRI, and Ultrasound, where the scanner captures cross-sectional slices along
 the z-axis. These slices are then stacked and reconstructed into a 3D volume. Other
methods include 3D laser scanning and structured-light imaging, which capture depth
information to form 3D data.
 Computed Tomography (CT) Scans
 Magnetic Resonance Imaging (MRI)
 Ultrasound (US) Imaging
 Micro-Computed Tomography (micro-CT)
 Laser Scanning/Triangulation
 Structured Light Scanning
 Photogrammetry / Structure from Motion (SfM)
[Link] is the use of slicing the data set?
Slicing a dataset in image processing involves dividing a larger dataset or individual images
into smaller, manageable subsets. This technique is used for various reasons related to model
training, evaluation, efficiency, and analysis. The primary use of dataset slicing in the context
of machine learning is to create reliable and generalizable models.
 Training, Validation, and Testing: Datasets are commonly sliced into separate training,
validation, and test sets.
o The training set is used to teach the model patterns and relationships within the data.

[Link] are the advantages and disadvantages of volumetric display?

Volumetric displays offer significant advantages in providing a true 3D viewing experience
with accurate depth perception, but they are limited by high costs, lower resolution compared
to 2D screens, and complex technical requirements.
Advantages of Volumetric Displays
 True 3D Visualization
 Realistic Depth Cues
 Multi-User Viewing
 Enhanced Understanding and Performance
 Natural Interaction
Disadvantages of Volumetric Displays
 High Cost
 Limited Resolution and Size
 High Computational Demands
 Occlusion Challenges
 Hardware Complexity and Maintenance
 [Link] to construct stereo views of image?

Constructing stereo views in image processing involves using two images captured from
slightly different horizontal viewpoints or, in the case of a single input image, generating a
second view by inferring depth information and applying a warping transformation. Stereo
views are constructed by generating two slightly different perspectives of the same scene—
one for the left eye and one for the right eye—to create a 3D perception.

1. Capture or Generate Two Viewpoints

2. Apply Geometric Transformation
3. Combine Left and Right Images into a Stereo Pair

This results in a stereo image that provides a 3D effect when viewed with appropriate display
techniques.

[Link] is the use of reflection of 3D images?

In image processing and computer graphics, 3D reflection (mirroring) is used to create
realistic visual effects, aid 3D reconstruction, and support various applications
in computer vision, medical imaging, and virtual reality.
Key Uses and Applications
 Photorealistic Rendering in Computer Graphics
 3D Reconstruction and Modeling
 Symmetry Detection and Analysis
 Medical Imaging and Education
 Virtual and Augmented Reality (VR/AR)
 Image Processing Operations
[Link] dull surface
In image processing, a dull surface generally refers to an object's appearance in an image
that exhibits low contrast, poor light reflection, or a lack of luster due to its physical
properties or poor image quality. The characteristics are related to how light interacts with the
surface and how the image sensor captures that information.
Key aspects in image processing include:
 Low Light Reflectivity/High Diffusion
 Lack of Contrast
 Texture and Micro-topography
 Subjective Perception
[Link] Shiny surface
 In image processing, a shiny surface is defined as a material that produces specular
reflection, where most of the incident light is reflected in a single, mirror-like direction. A
shiny surface in image processing refers to a surface that reflects a high amount of light,
producing specular highlights or bright spots in an image. These surfaces cause strong
directional reflections, making intensity values vary sharply and often creating glare or
saturation in captured images.

Key Characteristics in Image Processing

 Specular Reflection
 Highlights and Glare
 Image Formation
 Computational Considerations
[Link] about feature specific improvement
Feature-specific improvements in image processing involve techniques that enhance, extract,
or manipulate specific characteristics within an image, ranging from fundamental features
like edges, corners, and textures to high-level semantic features through deep learning.
These improvements are crucial for various applications, including medical diagnostics,
facial recognition, and remote sensing.
Key Feature Improvements and Techniques
1. Edges and Boundaries
 Canny Edge Detector
 Sobel and Prewitt Operators
 Morphological Operations
 2. Corners
 Harris Corner Detector
 Shi-Tomasi Corner Detector
 3. Texture and Shape
 Local Binary Patterns (LBP)
 Gabor Filters
 Shape Descriptors
[Link] the various sources of 3D data and how is the 3D image represented?
In image processing, 3D data is acquired from various sources and represented using several
methods, with the choice depending on the application and desired level of detail.
Sources of 3D Data
 Medical Imaging Scanners
 Laser Scanning and LiDAR
 Photogrammetry and Stereo Vision
 RGB-D Sensors
 Computer-Aided Design (CAD) Software
 Procedural Generation and Physical Simulation
 How the 3D Image is Represented
Once acquired, 3D data is typically represented using one of several common structures:
 Voxel Grids
 Point Clouds
 Meshes
 Implicit Surfaces
 Multi-view Images
[Link] are the reflection of 3D image visualization?
In 3D image visualization and processing, reflection refers to a geometric transformation that
produces a mirror image of an object or an entire scene across a specified plane. This process
is crucial for creating realistic rendering of reflective surfaces and can also be used as a
preprocessing operation in image analysis.
Key Aspects of 3D Reflection
 Mirror Image: A reflected object is a mirror image of the original object, with the same size
and shape, but located on the opposite side of the reflection plane.
 Reflection Plane: Unlike 2D reflection, which uses an axis, 3D reflection occurs across a
plane. The most common standard reflection planes are the XY, YZ, and XZ planes.
 Coordinate Transformation: The reflection process is achieved by transforming the
coordinates of each point of the object using a reflection matrix. For example, reflection
across the XY plane involves flipping the sign of the Z-coordinate (leaving X and Y
unchanged):
o Original coordinates: P(x, y, z)

o Reflected coordinates: P'(x, y, -z)

[Link] are the surfaces of 3D image visualization?

In 3D image visualization, "surfaces" refer to various methods of representing the boundaries
or outer appearance of 3D objects, rather than just raw volume data. These are used to create
visible models and typically fall into categories based on their mathematical or geometric
properties.
The primary types of surfaces and surface representations used in image processing and
computer graphics are:
 Polygon Surfaces (Mesh Modeling)
 Isosurfaces
 Spline Surfaces and NURBS
 Parametric Surfaces
 Implicit Surfaces
 Point Clouds
 Digital Surfaces/Faces
[Link] are the classifications of measurements on 3D images?
In image processing, measurements on 3D images (volumetric data or point clouds) can be
broadly classified based on the type of information extracted, such as geometric
measurements, intensity statistics, and topological features.
Geometric Measurements
These measurements quantify the physical characteristics of objects within the 3D space,
which are defined by X, Y, and Z coordinates.
 Dimensional measurements
 Volume
 Surface area
 Shape analysis
 Spatial position and alignment
Intensity Statistics
For 3D images represented as voxels (3D pixels) with intensity values (e.g., CT or MRI
scans), measurements often involve analyzing these intensity values within specific regions.
 Mean/median intensity
 Standard deviation/variance
 Histograms and intensity profiles
Topological Features
These refer to the properties that remain unchanged under continuous deformations and
describe the object's connectivity and structure.
 Connectivity
 Number of objects/holes
 Boundary representation
[Link] to determine the position of features in 3D images?
Determining the position of features in 3D images in image processing is typically achieved
using various computer vision and machine vision methods that leverage depth information.
 The primary techniques involve acquiring 3D data and then applying algorithms to pinpoint
feature coordinates in a 3D coordinate system (x, y, z).
Key Methods for 3D Feature Positioning
The approach used depends largely on the type of input data (single 2D image, stereo images,
LiDAR point cloud, etc.).
 Stereo Vision (Binocular Vision): This method uses two or more images of the same scene
taken from different viewpoints.
o Process:

o Feature Extraction

o Feature Matching

o Triangulation

 LiDAR and Laser Scanning: These methods are based on measuring the time it takes for a
laser pulse to return to the sensor.
o Process

 Structure from Motion (SfM)

 Monocular Vision Methods (Single Image)