Final Iteration Flashcards
(90 cards)
1
Q
What is the Nyquist rate?
A
The industry standard sampling rate which dictates that samples should be taken at a rate that is double the frequency of the highest signal.
2
Q
What is under-sampling?
A
Sampling at a rate below the Nyquist rate, also said as a sample rate that is too coarse
3
Q
How does Aliasing occur?
A
Results from under-sampling. Occurs when two signals are indistinguishable from each other when sampled.
4
Q
What is anti-aliasing?
A
Anti-aliasing - used when re-sampling. Makes sure to smooth out high frequency signals, so that it is ‘impossible’ to see the alias
5
Q
What is down-sampling?
A
Occurs when shrinking an image. Takes a local area of pixels, computes the average between them, and generates a new pixel using that value
6
Q
What is up-sampling?
A
Used when enlarging images. Takes a local area of pixels, and interpolates pixel values outside of that area, by gathering the average of those pixels, and passing that value through a fit function.
7
Q
What is quantisation?
A
A measure of light intensity. Determines grey level/colour resolution to be represented at each pixel
8
Q
What is re-quantisation?
A
For each pixel, divide its value by a constant, with the aim of reducing the number of grey levels/colours it can represent.
9
Q
How do you acquire colour images?
A
Either use an expensive camera with three CCDs, one for each primary colour of the RGB colour space, or use one CCD and a Bayer Pattern
10
Q
What is the Bayer Pattern?
A
A pattern of RGB colouring, where one colour is measured, typically from sampling, and the other two are estimated. One colour’s value is estimated by looking at neighbours (plus shape), whereas the other is measured using only diagonal colours.
11
Q
Why use greyspace over colour space?
A
Many techniques were developed with single value pixel images in mind
Reduces the amount of information in the image, which makes processing easier.
12
Q
How do you convert from RGB to greyscale?
A
Normal conversion - 0.3R + 0.59G + 0.11B
Green weighted - G-(R+B)/2
13
Q
What is an alternative colour space?
A
HSV:
Hue - the general colour
Saturation - the strength of the colour
Value - how light or dark the colour is
14
Q
How can you perform a linear transform?
A
Using addition and multiplication on a pixel’s value. More specifically - G(x,y) = a x f(x,y) + b
Where a is the gain, and b is the bias
15
Q
What does gain represent?
A
Gain is the level of contrast
16
Q
What does bias represent?
A
Bias is the level of brightness
17
Q
What is the dynamic range?
A
The range of values a pixel can represent e.g. [0, 255] means a pixel can be any value between 0 up to 255.
18
Q
What is contrast stretching?
A
Changes the dynamic range from s(min) and s(max) to t(min) and t(max) i.e. it changes the dynamic range’s minimum and maximum values
19
Q
What is an example of non-linear transformation?
A
Thresholding
20
Q
What is Grey Level Slicing?
A
Highlights a specific range of intensities, and can be used to preserve certain grey levels or reduce certain ones to the specified level.
21
Q
What is gamma correction?
A
Transforms an image so that it generates a voltage which will display the correct value of the pixel, rather than a value that is slightly off (which is normally what a computer monitor will do without GC)
22
Q
What is spatial filtering?
A
Spatial filtering - affects a small region, rather than an individual pixel (see linear filtering)
Contains more information, without sacrificing too much information about objects, edges, etc…
23
Q
What is the equation for image noise?
A
Recorded pixel value + a random noise value
24
Q
What is Gaussian noise?
A
Noise that exceeds a certain range of the Gaussian. Higher levels of variance will result in more noise.
25
How do you reduce noise?
Overlay multiple copies of the image on top of each other, and then produce an image where each pixel is the average across all the other pixel values in that location. Alternatively, applying filters can reduce noise e.g. mean filtering
26
What is convolution?
The process of applying a filter to an image
27
What is Gaussian filtering?
The process of applying a filter whose values are determined by a Gaussian function. Higher weight is given to pixels near the source pixel (origin).
28
What are discrete Gaussian filters?
Create a small square window which samples the Gaussian function, and normalises the results so that the filter entries add to 1.
29
How do you determine the Gaussian filter's size?
Depends upon the variance. A higher variance leads to more values being included that are above the 98% threshold.
30
What is a separable filter?
A 2D filter which can be split into two 1D filters e.g. a 2D Gaussian filter can be split into two 1D Gaussian filters, a horizontal one and a vertical one
31
How is Salt and Pepper noise generated?
When a faulty sensor registers either an error for a sample (black), or false saturation (white)
32
What is a Median filter?
Add up all of the pixel values, then divide by the amount of pixels to generate the median value. Then, apply that median value to pixels within it's radius.
33
What is anisotropic diffusion?
Anisotropic - not the same all sides
Diffusion - spreading out
Anisotropic Diffusion - making each pixel more like neighbouring pixels that it is already similar to.
34
What is the similarity function, and how does it work?
Calculated by using (D-d)/D, where D is the maximum possible difference, and d is the difference between the two target pixels.
S(p,q) is near to 1 - pixels are borderline identical
S(p,q) is near to 0 - pixels are almost entirely opposite to each other
S(p,q) means the new value at pixel p is based on all its neighbours, called q in this case.
35
What happens as the K value increases in Anisotropic diffusion?
Higher K value - greater smoothing, mostly preserves edges
36
What is Bilateral filtering?
Works by using two Gaussians. One weighs the value of pixels near the source pixel, whilst one weights the value of pixels similar to that of the target pixel.
37
What is Otsu Thresholding, and how does it work?
An adaptive thresholding technique
Assumes histograms are bimodal. Computes the weighted sum of the histogram, and selects the smallest T. This then becomes the new threshold for that area.
38
How does Unimodal Thresholding work?
Assumes histograms are unimodal.
Draws a line from the highest peak, down to the furthest bin's peak. The bin that is the furthest away from that line is set as the threshold.
39
What are connected components, and how are they calculated?
Turns a binary image into one that is labelled. Any areas that are connected via 1s is assigned a label. Anything with 0s is not labelled.
40
What is dilation and erosion, and how do both work?
Dilation - expands the foreground
Takes the structuring element's origin, and puts it 'on top' of the target pixel. Any background pixel that the structural element overlaps becomes a part of the foreground.
Erosion - Shrinks the foreground
Takes the structuring element's origin, and puts it 'on top' of the target pixel. If there are any background pixels within the range of the structural element used, then set the target pixel to the background.
41
What is opening and closing, and what are their results?
Opening - erode, then dilate
Smoothes contours, and eliminates protrusions
Closing - dilate, then erode
Smoothes contours, fuses small gaps, and fills in small holes
42
How can you edge detect using dilation/erosion?
Take the original image
Apply either erosion or dilation
Then, with the new image, subtract that from the original. The edges will now be displayed
43
How do you work out the 1st derivative?
Calculated by subtracting the current element from the one prior to it.
Also written as (x+1)-x
44
How do you work out the 2nd derivative?
(x+1) + (x-1) - 2(x)
I.e. element before current + element after current - 2(current element)
45
How is 1st derivative filtering modelled?
f is roughly |Gx|+|Gy|
46
What is the Gx and Gy for Roberts' Cross operator?
Gx:
1|0
0|-1
Gy:
0|1
-1|0
47
What is the Gx and Gy for Sobel Operator?
Gx:
-1|0|1
-2|0|2
-1|0|1
Gy:
-1|-2|-1
0|0|0
1|2|1
48
How do you produce a sharpened image (specifically focusing on edges)?
Take image
Apply Gaussian smoothing
Subtract the smoothed image from original to obtain unsharp mask
Add the unsharp mask to the original image
49
Why is 2nd derivative better for edge sharpening than 1st Derivative?
Produces a stronger reaction to fine details, and has a simple implementation.
50
What is a common edge sharpening filter and what is an advantage of using it?
Using a Laplacian filter
Advantage - simple implementation via convolution
51
What is the layout for a Laplacian filter both in normal form, and also for single operator form?
Normal:
0|1|0
1|-4|1
0|1|0
Single Operator:
0|-1|0
-1|-5|-1
0|-1|0
52
How do you find an edge with 1st derivatives?
Look for peaks in the 1st derivative, or zero-crossings (i.e. crossing the x-axis) in 2nd derivative
53
What are some properties of Roberts' Cross Operator?
Very efficient - uses only 4 pixels, and only subtracts and adds
Very susceptible to noise, and only reacts to very strong edges
54
How can you detect edges using 2nd derivative filters?
Using Marr-Hildreth - convolving the Laplacian of a Gaussian OR applying Gaussian smoothing, followed by a Laplacian
55
What is the difference between 1st derivative edge detection and 2nd derivative edge detection?
1st derivative peaks - strong response at edges, but also responds to noise. Peak detection and threshold selection need care
2nd derivative zero-crossings - well-defined, easy to detect. Edges must form smooth, connected contours, but tends to found off on corners.
56
What does a good edge detector require?
Good Detection
Good localisation
Minimal response
57
What is the Canny operator?
1st derivative of a Gaussian smoothed image.
Most implementations are 2D Gaussian smoothing + Roberts' style derivative
58
What is Non-Maximal Suppression?
Check if pixel is a local maximum along the gradient direction, and select a single maximum across the width of the edge.
59
What is Thresholding with Hysteresis?
Test each pixel independently. Industry standard allows a band of variation, but assumes continuous edges. User still selects parameters, but at the cost of less precision.
Idea is to keep weak edges connecting strong edges if the strong edges are exceptionally strong, and the weak edges are not exceptionally weak.
60
What is the formula for a histogram, and how does it change when it is a normalised histogram?
P(Rk) = Nk
Rk = kth grey level
Nk = amount of pixels at kth grey level
Normalised:
P(Rk) = Nk / N
N = amount of pixels in the image
61
What are some additional properties of normalised histograms?
A normalised histogram has bins that add up to 1.0
Each bin gives the probability of that grey level appearing in the image.
62
What does a low contrast and high contrast look like in histograms?
Low contrast - Very minimal spread, exceptionally high spikes
High Contrast - very minimal height, extreme spread across the histogram.
63
What does dark and light look like in a histogram?
Dark - Very left-hand side is full of extreme spikes
Light - very right-hand side is full of extreme spikes
64
What does histogram equalisation aim to do?
Aims to spread out the values across the histogram, so that the new image has a nearly uniform distribution of pixel values across the histogram i.e. no spikes at certain grey levels.
Increases contrast of overall image.
65
How do you equalise a histogram?
Take the R values and the Nk values. To calculate the Pr(Rk) values, you need to divide each Nk by the sum of all Nk values.
Next, using Pr(Rk), you can work out T(r) by taking the value for the first Pr(Rk), and copying it across. Then, for each subsequent value, add it, and any previous values, on top. That is their T(r) value.
Next, take the T(r) value, and multiply it by the maximum R value (should be a whole number).
Finally, for each of those values, round them to the nearest whole number (Note - always round down, no matter what)
Using these values, you can plot a new histogram using specifically Pr(Rk) and the rounded Sk values (the rounded numbers).
Design a histogram such that Pr(Rk) is the value on the y-axis, and the rounded value is on the x-axis. If there are multiple values for one rounded value, then add them together, which will result in the final y-axis point for that rounded value.
66
What are some strengths and weaknesses for histogram equalisation?
Works well when the input isn't too noisy, and there aren't any exceptionally bright or dark areas that could overpower the new image.
A way to counter these problems is using equalisation of histograms in local areas of the image.
67
What are some problems with colour histograms?
Noise and/or different camera responses can give similar images with very different histograms
Histogram resolution - may need many bins to store all the colours, which can get very expensive very quickly (expensive in terms of storage and memory cost)
The illumination may be coloured - same object may generate a different histogram under different lighting
They ignore spatial information
68
How do you calculate the histogram intersection?
Using two 'arrays' for histograms, put one above the other. Then, for each pair (one above, one below), take the smallest value. Add all those values together and you get the histogram intersection score. A higher score means more intersection.
69
What different types of image matching with colour histograms are there?
Textual query - finds images in the database based on key words associated with each image - can get expensive
Content-based retrieval - use the shapes, textures and colours within the query image to search for images that have similar properties.
70
How does region/object-based query work?
Divide the image into small windows, and see how much of the target colour is present, by highlighting those pixels in the image that are similar to those in the query.
71
What different types of segmentation are there and give a brief explanation of what each one does?
Clustering - seeks groups of similar pixels, with no regard for where they are, and views images as uncorrelated data.
Region-based - starts with a 'seed' (origin pixel), and computes a similarity value for comparison with other neighbours e.g. average grey level value. If the neighbour is close to that value, adds it to the region.
Split and merge - Splits up the image into many regions, then merges them based on their similarity to each other.
72
What are quadtrees?
Trees which represent how the image has been split - used within split and merge. Each time it splits the image/region, it splits it into 4 parts, hence quad
73
How does watershedding work?
Edges are represented as ridges in an otherwise massive 'valley' where basins are regions in the image. Figuratively, it raises the water level, and when the water overflows a ridge, it detects an edge. In practice, it orders pixels from lowest value to highest, and goes through them one by one, assigning labels to them. if it doesn't have a label, and no neighbours have a label, then give it one. If it has a neighbour with a label, then give it that one. If it has two or more neighbours with different labels, then mark that pixel. It is considered as an edge.
74
How does SLIC work?
Initialise cluster centres on pixel grid in S steps - image has N pixels and you want K superpixels, where each superpixel is roughly a square region of N/K pixels. Thus, S = square root of (N/K)
Move centres to the position in a 3x3 window with the smallest intensity gradient - moving centres away from edges
Compare each pixel to all cluster centres with 2S pixels and assign it to the best matching cluser
Recompute cluster centres as mean colour and position of the pixels belonging to each cluster
Repeat steps 3 and 4 until total change made to position and colour of centres is below a threshold, or for a fixed number of iterations.
75
What is Template Matching?
Take a straight line, and match it to all possible image orientations and positions, and compute a measure of fit to the edge data - incredibly expensive, don't do this :(
76
What is the Hough Transform?
Takes lots of edges, transforms them into lines that exist in m,c space and finds lots of places where they intersect and returns those parameter points
77
What is smoothing, sharpening and band limiting in regards to Frequency Domain Processing?
Smoothing - low pass filtering - attenuate high frequency components
Sharpening - high pass filtering - attenuate low frequency components
Band limiting - set all components to 0 outside a given frequency range
78
What is VLE?
Variable Length Encoding
Assigns fewer bits to more commonly used grey levels than to less probable ones - in order to save space
79
What is Fidelity Criteria?
A measure of success of psychovisual redundancy compression. Lower score is better quality, and score is determined by comparing the original image to the newly compressed version of the image.
80
What is the function of a Mapper?
Transforms input data in a way that facilitates reduction of interpixel redundancies - reversible
81
What is the function of a Quantiser?
Transforms input data in a way that facilitates reduction of psychovisual redundancies - not reversible
82
What is the function of a Symbol Coder?
Assigns the shortest code to the most frequently occurring output values - reversible
83
What is Entropy?
The average information content of an image, a measure of the histogram dispersion.
84
How does GIF compression work?
Map vector values (R, G, B) onto scalar values.
Multiple vectors map to each scalar
For each pixel in the original image, find the closest colour in the Colour Table. Record the index for that colour. To reconstruct the image, place the indexed colour from the Colour Table at the corresponding spatial location.
85
How do you build a palette?
Find clusters of pixels that are close/similar in colour, and combine them all to form a 'generic' colour, which represents them all. Replace them all by this single colour.
86
What is Differential Pulse-Code modulation?
Code the difference between adjacent pixels.
Prediction is that the next pixel value is equal to the current one, and you need the first value to provide a point of reference. It is lossless and is of a lower entropy.
87
What is the step by step process for converting an image to a JPEG?
Break into 8x8 blocks, and start from the top left and work your way through as if you were reading a book. For each one, subtract half the maximum intensity value. Once finished, apply the 2D-DCT to this block.
88
What are AC and DC components?
DC components summarise patch intensity
AC components are quantised - divide the DCT block by values in a quantisation table, with different tables for luminance and chrominance
89
What are some properties of histograms in general?
Invariant to translation and rotation
Change slowly as viewing direction changes
Change slowly with object size
Change slowly with occlusion
90
List the three different types of redundancy?
Spatial
Psychovisual
Coding
