Random

Question 1

What would be a reasonable sampling frequency for a signal?

Answer 1

Shannon’s theorem suggests a sampling of at least 2xf. However, for a better reconstruction of the signal, a reasonable sampling frequency should be 10 x f

Question 2

Imagine we want to build a bionic eye that adds X-ray vision to a human and makes it possible to see through the walls. Identify two ethical questions related to this system. [5p]

Answer 2

Imagine we want to build a bionic eye that adds X-ray vision to a human and makes it possible to see through the walls. Identify two ethical questions related to this system. [5p]

Question 3

What is an ultrasound sensor? What does it measure, and how does it work?

Answer 3

An ultrasound sensor, also known as sonar, makes use of echolocation such that it sends out a beam with a frequency beyond human hearing in order to measure the time needed to receive the echo reflected by the object.

Question 4

What is a passive opto-sensor? What does it measure?

Answer 4

A passive opto-sensor would sense and measure the ambient light only.

Question 5

What is an active opto-sensor? What does it measure, and how?

Answer 5

An active opto-sensor would produce a red beam and measure the intensity of light reflected back from the illuminated surface. This also allows it to make a distinction between coloured surfaces.

Question 6

What is the frequency of a sinusoidal signal with a period of 5 sec?

Answer 6

The formula for calculating the frequency of a sinusoidal signal is f = 1/5. Thus, its frequency is 1/5 Hz.

Question 7

What is the FFT? Sketch FFT algorithm as a black box, and identify its inputs and outputs.

Answer 7

Fast Fourier Transform (FFT) is a fast and efficient algorithm for Discrete Fourier Transform (DFT), an algorithm to transform a time-domain representation of a digital signal into a frequency-domain signal.

Its inputs are a vector containing the digitised signal samples (x(n)) and the number of samples (N) with an output of a vector containing some amplitudes, ak.

Question 8

We applied the FFT to a sampled signal with 1000 samples and sampling frequency Fs=8000Hz. What does the vector returned by the FFT contain?

What are the first and last elements? What is the distance between 2 adjacent elements?

Answer 8

The vector returned by FFT contains amplitudes from a0 to a1000.

Each frequency corresponding to each amplitude is calculated by the formula: fk = kFs/N. Thus, for a0 = f0 = 08000/1000 = 0 and a1000 = f1000 = 1000*8000/1000 = 1000

Question 9

Enumerate the steps in a classification process

Answer 9

· Sense data from the environment.
· Preprocess sensor data.
· Explore data. Define the classes.
· Identify possible attributes (features) which will allow discrimination between classes. They should be compact and have discriminatory power.
· Process data to extract these features.
· Design a classifier and train it. For each training example, provide at least one feature vector.
· Evaluate the classifier. The classifier should be assessed to see how well it can generalize to new examples.

Question 10

What is a rule-based classifier?

Answer 10

A rule-based classifier distinguishes between different classes by using a collection of this kind of "if...then" decision rules. The general form for a rule is as follows:
(Condition) ! y,
where Condition is a conjunction of features, and y is the class label.

Question 11

Define the concept of frequency spectrum.

Answer 11

Frequency spectrum analysis aims to identify which frequencies are present in a time-varying signal, at a given moment.

Question 12

What are the three waves of computing according to Mark Weiser?

Answer 12

Mainframe, personal computing, ubicomp

Question 13

Explain Mark Weiser’s vision of Ubiquitous Computing

Answer 13

He envisioned a world where computing and applications are embedded in everyday objects like cars, televisions and clothes.

Question 14

What does MEMS mean? Give some examples.

Answer 14

Microelectro-mechanical systems are a promising technology which is constantly evolving.

They are microns to millimeter-scale devices that integrate mechanical mechanisms together with the necessary electronic circuits, on a single silicon chip. MEMS can perform two basic functions: sensing and actuation. Examples of MEMS sensors are the accelerometers used to release safety airbags in modern cars.

Question 15

How is a map represented in a computer-based navigation system?

Answer 15

In a digital road map, the roads are divided into segments, and each road segment is represented by two edges in the graph. The cost associated with each edge is not per definition the distance between the nodes, as you might think, but can be the expected travel time, the amount of money, fuel consumption, or number of traffic lights encountered on the way.

From the graph-theoretical point of view, the route planning in a navigation system is defined as a path-finding problem in a graph. Planning the optimal route from A to B is equivalent to finding a shortest path between the two nodes. Shortest means in this case a path with the lowest possible cost.

The most well-known algorithm for computing shortest paths is the Dijkstra’s algorithm. This algorithm works by visiting all nodes in the graph, beginning at the starting point. The search expands outward until it reaches the goal

Question 16

What is an actuator? Give a few examples

Answer 16

Actuators are transducers that receive the electrical signal from the controller and use it to affect back the environment. Examples of actuators are heaters, lamps, LEDs and displays, speakers, and motors

Question 17

What is thresholding? Where do we need it? How does it work?

Answer 17

Thresholding is another, very useful, point-processing operation applicable to grayscale images. It transforms a grayscale image into a black and white (binary) one, by choosing a gray level T in the original image and then turning every pixel black or white, according to whether its grey value is greater than or less than T - the threshold level.

In many image processing systems, thresholding is a key step to a segmentation (separation) of the foreground (information) from the background (noise). By selecting a “good” threshold value between the two histogram peaks, one can successfully separate the foreground objects from the background.

Segmentation by thresholding gives the best results if the image has a bimodal histogram. Bimodal means that the histogram has two prominent peaks, where one peak corresponds to the background pixels and the other peak to the foreground pixels.

Question 18

What is the big disadvantage of filtering?

Answer 18

there is no perfect filter, appropriate for all images. You should always try to apply different filters, with different sizes, and see which one has the best effects. In any case, be aware that filtering has the same effect as chemotherapy in cancer treatment. Since the processing applies to all pixels in the image, not only the noisy ones, there will be always a price to be paid; the image might become better, but not the same anymore.

Question 19

What is a structuring element(STREL) ? Give some examples

Answer 19

The kernel in a morphological operation is called structuring element (STREL) and contains only “0”s and “1”s. Normally, the pattern of “1”s forms a box or a disk. Structuring elements have a designated center pixel.

The art of morphological processing is to choose the right shape of the structuring element that suits the particular application. In general, a box-shaped structuring element tends to preserve sharp object corners, whereas a disk-shaped structuring element tends to round the corners of the objects.

Morphological operators take a binary image and a structuring element as input and combine them using a set operator (intersection, union, inclusion, or complement).

However, a STREL is not applied in the same way as the kernel in correlation. Other operations (called Fit and Hit) are used instead of using multiplications and additions.

In a particular situation, you should always try to adjust the size of the structuring element until the best effect is obtained.

Question 20

What is a frequency response graph?

Answer 20

A filter’s frequency-selective behaviour is specified by its frequency response, a function that can be visualized in a plot, with frequency on x-axis and amplitudes on y-axis

In these graphs, we can distinguish the following characteristics of a filter: the passband that refers to those frequencies that are passed, the stopband that specifies those frequencies that are blocked, and the transition band, that lays in between. The division between the passband and transition band is called the cutoff frequency.

Question 21

What characterizes a “good” feature for classification? Give an example

Answer 21

Features have discrimatory power, thus a good feature is one which has sufficient discriminatory power in order to allows us to successfully solve the coin classification problem.

Such a feature can be frequency hz, diameter size, etc.

Question 22

What typical features can be used for image recognition?

Answer 22

Image recognition is widely used in pervasive computing systems. Typical features used for image pattern recognition include single parameter shape descriptors, statistical image movements, and zoning.

However, perfect, universal set of features do not exist, these are simply different approaches with their advantages and disadvantages.

Question 23

What typical features can be used for speech recognition?

Answer 23

An Automatic Speech Recognition (ASR) module is an inherent component in smartphones, voice-to-text translators in medical world, and live subtitling systems for TV. Its goal is to accurately and efficiently convert a speech signal into a text transcription, independently of the device used to record the speech, the speaker accent, or the acoustic environment in which the speaker is located.

Thus, suitable, advanced features for sound and speech recognition are frequency spectrum and mel-freuquency cepstral coefficients (MFCC).

Question 24

What does learning mean?

Answer 24

Learning is defined as getting better at some task through practice.

Machine learning algorithms, which modify and adapt their actions so that their actions become more accurate, implement different types of learning, such as supervised, unsupervised, reinforced, and evolutionary.

Question 25

What is a perceptron? Can a perceptron classify? How?

Answer 25

The simplest neural network is a single-layer one, with one layer of inputs, and one layer of outputs, called a (single-layer) perceptron.

The advantage of such a neural network is that it can learn how to classify when presented with inputs, by coming up with classes they belong to, codified with 1 or 0. However, a perceptron can only discriminate between two classes, but with more than one output neuron, more classes can be labelled.

Learning is done by adjusting these weight coefficients, during training. A training set contains examples of inputs (the features x1, x2, x3) and the corresponding correct output (the class c = 0 or 1). For each pair (inputs, output) from the training set, the perceptron will start a learning process.

Question 26

Why is completely automatic supervised learning not yet possible?

Answer 26

Although a lot of steps can be performed automatically, human experts are still essential to specify which classes we have to consider, which features to use and to train the classifier for supervised classification.

Question 27

Although a lot of steps can be performed automatically, human experts are still essential to specify which classes we have to consider, which features to use and to train the classifier for supervised classification.

Answer 27

A solution for dealing with uncertainty in automatic classification is to use probabilistic models.

These are algorithms that improve themselves through experience, for example, naïve Bayes classifiers and Hidden Markov Models (HMM). All these models use a very important theorem, known as the Bayes rule.

Question 28

Define the terms prior probability and conditional probability. Give an example for each.

Answer 28

The Bayes Rule makes use of A and B, random events, where P(A|B) is the conditional probability, meaning that “A is true if B is true”, P(A) is the prior probability that A is true, P(B) is changed depending on A, resulting in calculating P(B|A).

Question 29

explain how colours are represented in an RGB digital image

Answer 29

RBG images represent colors as 3 overlapping matrecies where the colors shown using a vector which
calculates the true color using the the respective Red Green and Blue value at each position and layer
in the array