Image analysis

Question 1

What is the objective of image segmentation

Answer 1

Splitting the image into meaningful parts/ regions belonging together

Question 2

Describe mean shift segmentation

Answer 2

1. Create a map of the number of pixels at each color in a color space.
2. Do a gradient ascent on each pixel in this map.
3. Assign pixels on the same peak to the same segment

Question 3

What do active contour algorithms do?

Answer 3

They fit curves to object boundaries

Question 4

Name some split and merge algorithms

Answer 4

1. Watershed
2. Region splitting
3. Region merging
4. Graph-based algorithms

Question 5

What operations are performed in a morphological opening?

Answer 5

Erosion then dilation

Question 6

Which operations are performed in a morphological closing?

Answer 6

Dilation then erosion

Question 7

What properties are important for features used for classification

Answer 7

1. They should contain information that will aid in discriminating between classes.
2. They should exhibit low in class variance and high between class variance
3. They should not contain redundant information.

Question 8

What is the difference between first order and second order statistics in Histogram features

Answer 8

First order statistics is the gray-level distribution, second order statistics contains spacial information.

Question 9

Name some problems where Image pattern recognition can be used

Answer 9

1. Face detection
2. License plate/bar code detection
3. Medical imagining
4. Fingerprint recognition
5. Character recognition

Question 10

Name som types of clasifiers

Answer 10

1. Neural Nets
2. Bayes classifiers
3. Support Vector Machines (SVM)
   4- K-Nearest neighbor
   …

Question 11

Briefly explain Otsu’s method

Answer 11

Otsus method is a thresholding segmentation algorithm. The thresholds are adapted to minimize in-class variance and maximise between class variance.

Question 12

Name some features we can extract from images that for example can be used for object classification.

Answer 12

• Colour features
• Gray level features
• Shape features
• Histogram (texture) features

Question 13

What do we mean by “the curse of dimensionality”?

Answer 13

1. With a high input dimensionality, the algorithm will easily overfit to the training data and not generalize well.

Question 14

What do we mean by “machine learning”? Give examples of some common methods.

Answer 14

Machine learning is a set of algorithms that adapt parameters in mathematical models to fit a set of data. Some common methods are Neural nets (MLP, convolutional..), k-means, Support vector machines, evolutionary algorithms…

Question 15

What do we mean by supervised learning

Answer 15

Supervised learning algorithms are algorithms that have training data with “targets”. The targets are the true solutions to the training data input. Neural nets are supervised learning algorithms.

Question 16

How can image descriptors be used for object detection?

Answer 16

We can train an algorithm to score the descriptors. We could, for example, use an SVM to give a score to the descriptors for how likely it is that the descriptor is from an image patch with a pedestrian.