w5 gemini

Question 1

What is the primary goal of image segmentation in computer vision?

Answer 1

To group together elements of an image that ‘belong together’ and segment these elements from all others.

Question 2

Differentiate between ‘grouping’ and ‘segmenting’ in the context of image analysis.

Answer 2

Grouping refers to bringing together elements that are similar or connected, while segmenting involves isolating these grouped elements from the rest of the image.

Question 3

List four main categories of methods used for image segmentation.

Answer 3

Thresholding, Region-based methods, Clustering, and Fitting.

Question 4

Explain the concept of ‘Feature Space’ in the context of image segmentation.

Answer 4

Feature space is a conceptual space where each element of an image (e.g., a pixel) is represented as a point based on its feature values. Similarity between elements is determined by the distance between their points in this space.

Question 5

Name at least three features that are commonly used in computer vision for determining if elements ‘belong together’.

Answer 5

Location (proximity), Colour/Intensity, Texture, Depth, Motion, Contour, and CNN features.

Question 6

Explain the role of ‘Similarity Measures’ in feature space for image segmentation.

Answer 6

Similarity measures quantify the resemblance between image elements based on their feature vectors. This helps in grouping elements that are considered alike.

Question 7

Provide an example of a similarity measure and briefly explain how it works.

Answer 7

Euclidean distance is a common similarity measure. It calculates the straight-line distance between the feature vectors of two elements in feature space; smaller distances indicate higher similarity.

Question 8

Why is ‘feature scaling’ important when using multiple features for image segmentation?

Answer 8

Features with different scales can disproportionately influence distance calculations. Feature scaling ensures that all features contribute equally to the similarity assessment.

Question 9

Explain the basic principle of ‘Thresholding’ for image segmentation.

Answer 9

Thresholding segments an image by setting a threshold value. Pixels with intensity values above the threshold are assigned to one group, and those below are assigned to another.

Question 10

What is the limitation of basic thresholding methods, and how can ‘local thresholding’ address this?

Answer 10

Basic thresholding uses a single threshold for the entire image, which can be ineffective with uneven illumination. Local thresholding applies different thresholds to different regions of the image to adapt to varying lighting conditions.

Question 11

Describe the ‘Hysteresis thresholding’ technique and its advantage over single-threshold methods.

Answer 11

Hysteresis thresholding uses two thresholds. Pixels above the high threshold are classified as foreground, and those below the low threshold are background. Pixels between the thresholds are classified based on their connectivity to foreground pixels, reducing noisy segmentation.

Question 12

What are ‘Morphological Operators’, and how are they used in conjunction with thresholding?

Answer 12

Morphological operators (like dilation and erosion) are used to clean up the results of thresholding by filling gaps, removing noise, and smoothing boundaries in binary images.

Question 13

Explain the ‘Dilation’ operation in morphological image processing.

Answer 13

Dilation expands the area of foreground pixels in a binary image. Background pixels adjacent to foreground pixels are changed to foreground.

Question 14

Explain the ‘Erosion’ operation in morphological image processing.

Answer 14

Erosion shrinks the area of foreground pixels in a binary image. Foreground pixels adjacent to background pixels are changed to background.

Question 15

What is the fundamental drawback of thresholding that ‘Region-based’ methods aim to address?

Answer 15

Thresholding does not explicitly consider the spatial relationships between pixels.

Question 16

Describe the ‘Region Growing’ segmentation method.

Answer 16

• starts with 1 seed pixel
• iteratively adds neighbouring pixels to the region based on a similarity criterion.
• continue until region stops growing
• pick another random seed (from those not assigned to a region) and repeat until all labelled

Question 17

Explain the ‘Region Merging’ segmentation method.

Answer 17

• starts with each pixel (or small regions) as a separate region
• iteratively merges adjacent regions based on their similarity
• continue merging adjacent regions until can’t anymore, mark region complete
• repeat on remaining until all finals

Question 18

Explain the ‘Region Splitting and Merging’ segmentation method.

Answer 18

Splitting;
1) begin by splitting image into quadrants
2) for each region- if all pixels are not similar, split into 4 sub quadrants
Continue until each region is homogeneous

Merging;
For each region - merge with similar neighbours
Continue until no more regions can merge

Question 19

What is a potential problem with ‘Region Merging’ related to the order of merging?

Answer 19

The order in which regions are merged can affect the final segmentation, especially if the properties used for merging are averages of the constituent parts.

Question 20

What are some general problems associated with region-based segmentation methods?

Answer 20

Meaningful regions may not have uniform properties (e.g., due to lighting variations), and it’s unusual for natural images to be composed of perfectly uniform regions.

Question 21

What is the core idea behind ‘Clustering-based Segmentation’?

Answer 21

Clustering aims to group image elements into clusters based on the similarity of their feature vectors, without prior knowledge of the number or characteristics of the clusters.

Question 22

Differentiate between ‘Partitional Clustering’ and ‘Hierarchical Clustering’.

Answer 22

Partitional clustering divides data into non-overlapping subsets, while hierarchical clustering creates a nested hierarchy of clusters.

Question 23

Explain the ‘K-means Clustering’ algorithm.

Answer 23

K-means randomly initializes cluster centers, assigns each data point to the closest center, recalculates the centers based on the assigned points, and repeats until the centers stabilize.

Question 24

What is a limitation of the K-means algorithm regarding the initial placement of cluster centers?

Answer 24

The final clustering result in K-means can be sensitive to the initial random placement of cluster centers, potentially leading to different outcomes for different runs.

Question 25

Describe a scenario where K-means clustering might perform poorly.

Answer 25

K-means performs poorly when clusters have significantly different sizes, non-globular shapes, or varying densities.

Question 26

Explain the basic principle of ‘Agglomerative Clustering’.

Answer 26

Agglomerative clustering starts with each data point as a separate cluster and iteratively merges the two most similar clusters until a stopping condition is met.

Question 27

Name three different ways to define ‘Inter-Cluster Similarity’ in Agglomerative Clustering.

Answer 27

Single-link (shortest distance), Complete-link (longest distance), Group-average (average distance), and Centroid (distance between means).

Question 28

Explain how ‘Segmentation by Graph Cutting’ works.

Answer 28

Graph cutting represents the image as a graph where nodes are image elements and edges represent similarity. Segmentation is achieved by cutting edges to partition the graph into subgraphs with minimum inter-subgraph similarity and maximum intra-subgraph similarity.

Question 29

What is the goal of ‘Normalized Cuts (Ncuts)’ in graph-based segmentation, and how does it address a limitation of basic graph cut methods?

Answer 29

Ncuts aims to overcome the bias of standard graph cuts towards small segments by normalizing the cut cost by the total strength of connections within each segment.

Question 30

What is the main idea behind ‘Fitting Methods’ for image segmentation?

Answer 30

Fitting methods try to represent a set of image elements using a mathematical model, and elements that fit the model are grouped together.

Question 31

Explain the ‘Hough Transform’ and how it can be used to detect straight lines in an image.

Answer 31

The Hough Transform maps points in image space to curves in parameter space. Intersections of these curves in parameter space correspond to lines that pass through multiple points in the image.

Question 32

Briefly describe how the ‘Generalised Hough Transform’ extends the basic Hough Transform.

Answer 32

The Generalised Hough Transform can detect arbitrary shapes, even those not easily expressed parametrically, by using a reference point and a lookup table describing the shape’s contour.

Question 33

What are some advantages of the Hough Transform?

Answer 33

Tolerance to gaps in edges
tolerance to occlusion
relatively unaffected by noise
detect multiple occurrences of a shape.

Question 34

What are some disadvantages of the Hough Transform?

Answer 34

Expensive in terms of memory and computation
potential loss of localization information
difficulty in locating peaks in noisy conditions
challenges in determining appropriate accumulator quantization.

Question 35

Explain the concept of ‘Active Contours’ (snakes) for image segmentation.

Answer 35

Active contours are spline curves that deform under the influence of internal forces (smoothness, bending) and external forces (attraction to image features) to minimize an energy function and fit object boundaries.

Question 36

What are the two main types of energy that influence the deformation of an active contour?

Answer 36

Internal energy (based on the shape of the snake) and external energy (based on the snake’s proximity to image features).

Question 37

What are some limitations of ‘Active Contours’?

Answer 37

Only works for smooth, closed contours, extremely sensitive to parameter settings, convergence is dependent on initial position, and no external force acts on points far from intensity gradients.

Question 38

Summarize the difference between ‘edge-based’ and ‘region-based’ segmentation methods.

Answer 38

Edge-based methods focus on finding discontinuities in image features to define boundaries, while region-based methods focus on grouping similar pixels based on their feature values.

Question 39

Summarize the difference between ‘model-based’ and ‘model-free’ segmentation methods.

Answer 39

Model-based methods attempt to fit the image data to a predefined model, while model-free methods rely purely on the properties of the image features without assuming a specific model.

Question 40

Explain the ‘top-down’ approach to segmentation.

Answer 40

Top-down segmentation uses prior knowledge or a model of the object to guide the segmentation process.

Question 41

Explain the ‘bottom-up’ approach to segmentation.

Answer 41

Bottom-up segmentation groups image elements based on their inherent similarities and spatial coherence without prior knowledge of the objects.