Lecture 3 - Feature Extraction and PCA

Question 1

What is feature extraction in the context of image analysis?

Answer 1

Feature extraction involves identifying and describing patterns or objects within an image, often using a vector of values to represent various aspects like shape, texture, or color.

Question 2

Explain the difference between global and local features.

Answer 2

Global features describe the entire image with a single vector, whereas local features describe specific patches within the image.

Question 3

What are the primary applications of feature extraction?

Answer 3

Feature extraction is used for tracking, registration, segmentation, 3D correspondence, retrieval, and recognition.

Question 4

Define Principal Component Analysis (PCA).

Answer 4

PCA is a statistical method used to reduce the dimensionality of data by transforming it into a set of uncorrelated variables called principal components, which capture the most variance in the data.

Question 5

What are the key challenges in feature extraction?

Answer 5

Key challenges include variations in viewpoint, illumination, background, occlusion, and intra-class variation.

Question 6

Describe the concept of “interest points” in local feature extraction.

Answer 6

Interest points are distinct and localizable patches within an image that can be reliably detected under various conditions. Examples include edges, corners, and blobs.

Question 7

Explain the importance of the structure tensor in corner detection.

Answer 7

The structure tensor captures intensity variations in different directions around a point, helping to identify corners where these variations are significant.

Question 8

What is the Harris corner detector?

Answer 8

The Harris corner detector is an algorithm that identifies corners by examining the eigenvalues of the structure tensor, looking for regions with high variations in intensity in multiple directions.

Question 9

Define “blob detection” and its significance.

Answer 9

Blob detection identifies areas in an image where some properties are roughly constant and differ from their surroundings, useful for detecting regions like cells in biomedical images.

Question 10

Describe the Scale-Invariant Feature Transform (SIFT).

Answer 10

SIFT is a feature extraction method that detects keypoints in an image and generates descriptors that are invariant to scale, rotation, and partially invariant to illumination and affine transformation.

Question 11

What is the role of Gaussian smoothing in edge detection?

Answer 11

Gaussian smoothing reduces noise by averaging pixel values with a Gaussian-weighted kernel, making edge detection more robust against noise.

Question 12

Explain the concept of non-maximum suppression in edge detection.

Answer 12

Non-maximum suppression is a technique used to thin edges by keeping only the local maxima in the gradient magnitude image, resulting in one-pixel-wide edges.

Question 13

Write the formula for the structure tensor used in corner detection.

Answer 13

Question 14

Provide the formula for the Harris corner response.

Answer 14

Question 15

Write the formula for the gradient magnitude used in edge detection.

Answer 15

Question 16

What is the eigenvalue decomposition of the covariance matrix in PCA?

Answer 16

Question 17

How does PCA reduce the dimensionality of data?

Answer 17

PCA reduces dimensionality by transforming data into a set of orthogonal components that capture the most variance, allowing the data to be represented with fewer dimensions.

Question 18

What is the main advantage of using local features over global features?

Answer 18

Local features are more robust to variations in viewpoint, illumination, and occlusion, as they focus on specific parts of the image rather than the entire image.

Question 19

Explain the significance of eigenvalues and eigenvectors in PCA.

Answer 19

Eigenvalues represent the amount of variance captured by each principal component, while eigenvectors define the direction of these components in the feature space.

Question 20

Describe the process of template matching in appearance-based methods.

Answer 20

Template matching involves shifting a template over an image and comparing the template to the image using similarity measures like sum-of-squared differences (SSD) or normalized cross-correlation (NCC).

Question 21

What are the limitations of using a single template for object recognition?

Answer 21

A single template may not account for variations in object appearance, requiring multiple templates to handle different views, scales, and lighting conditions.

Question 22

How does the Sobel operator approximate image gradients?

Answer 22

The Sobel operator uses convolution with specific kernel masks to compute approximate horizontal and vertical derivatives, which are then combined to obtain the gradient magnitude and direction.

Question 23

What is the role of the Laplacian of Gaussian (LoG) in blob detection?

Answer 23

The LoG is used to detect blobs by identifying extrema in the scale-space representation of the image, indicating areas with significant intensity changes.

Question 24

How does non-maximum suppression improve edge detection results?

Answer 24

Non-maximum suppression refines edge detection results by retaining only the local maxima in the gradient magnitude, leading to thinner and more accurate edges.

Question 25

Explain the concept of “invariance” in the context of feature descriptors.

Answer 25

Invariance refers to the property of feature descriptors to remain stable under various transformations, such as rotation, scaling, and changes in illumination, ensuring reliable matching across different views.

Question 26

Describe the process of SIFT feature extraction.

Answer 26

SIFT feature extraction involves detecting keypoints using the Difference of Gaussian (DoG) method, assigning a dominant orientation, and computing a descriptor based on local gradient orientations around each keypoint.

Question 27

What are the benefits of using PCA for face recognition?

Answer 27

PCA reduces the dimensionality of face images, capturing the most significant variations and allowing for efficient and accurate recognition by comparing principal components.

Question 28

How does the eigenface method utilize PCA for face recognition?

Answer 28

The eigenface method applies PCA to a set of training face images to extract principal components (eigenfaces), which are then used to represent and compare new face images for recognition.