Whole Image Methods

Question 1

What are histograms, and what are their uses?

Answer 1

The histogram of a digital image with grey levels in the range [0, L-1] is a discrete function p(Rk) = Nk
Where:
k = 0,1,…, L-1
Rk is the kth grey level
Nk is the number of pixels in the image with that grey level information
Histograms provide useful global information about the image, ease computation of some image properties, and can be manipulated to improve the image

Question 2

What is the difference between a histogram and a normalised histogram?

Answer 2

p(Rk) = Nk
Becomes
p(Rk) = Nk/n
where
n - width x height i.e. the total number of pixels in the image
The bins in a normalised histogram sum to 1.0, and each bin gives the probability of the corresponding grey level appearing in the image
The probabilistic interpretation is valuable in contrast enhancement and automatic thresholding

Question 3

How is a histogram represented for light and for dark images?

Answer 3

Dark images - Quite large spikes on the left hand side of the diagram. Nothing on the right as there is no light
Light images - Quite large spikes on the right hand side of the diagram. Nothing on the left as there is not much darkness

Question 4

How is a histogram represented for low contrast and high contrast?

Answer 4

Low contrast - huge spikes in the middle of the diagram
High contrast - lots of very small spikes evenly spread across the whole diagram.

Question 5

What is the aim in histogram equalisation?

Answer 5

Goal is to improve the contrast of an image:
To transform an image in such a way that the transformed image has a nearly uniform distribution of pixel values
This is more general than linear or piecewise contrast stretching - no parameters to specify ‘apply a transformation that makes the output histogram flat’

Question 6

How do histogram transforms work (histogram equalisation)?

Answer 6

Map an input histogram, r, onto a new histogram, s.
Assume r has been normalised to the interval [0,1] with r = representing black and r = 1 representing white:
s = T(r) where 0<=r<=1

Question 7

How do you perform histogram equalisation?

Answer 7

Compute the CDF (cumulative distribution function) of the input image
For each pixel in the input image, the corresponding output pixel intensity is calculated by using the CDF as a look-up table
CDF values will be in the range [0,1], scale the equalised image to fit the rang supported by the output image format
The histogram of the output image will be approximately uniform

Question 8

What are the strengths for using histogram equalisation?

Answer 8

Works very well when the input images:
Don’t have large bright or dark areas
Aren’t too noisy
If this is the case, then the results can be a bit skewed.

Question 9

What is a way to combat the weaknesses of histogram equalisation?

Answer 9

Using local equalisation - transforming a histogram for only a segment of the image, rather than the whole thing all at once

Question 10

How does histogram equalisation affect most histograms?

Answer 10

Usually shrinks the peaks, and strengthens the dips so that the overall average is better, and the angle of the curves on a line traced above all the points is smaller.

Question 11

What are text-based approaches when trying to find images in a database?

Answer 11

Retrieval of images is via text search
This is done by adding relevant words/phrases to each image
However, annotation is subjective, laborious and unnecessary

Question 12

What does context-based retrieval mean and what does it contain?

Answer 12

Indexes the image database on visual features, such as colour, shape and texture.
Queries are expressed in those terms or via visual examples
Simple approaches compute metric distance between the query image and each image in the database

Question 13

What are colour histograms and how are they created?

Answer 13

Choose a colour space - RGB, HSV
Divide the axes to create a reasonable number of divisions - trade-off detail for memory/computational cost
Build a histogram
Normalise if images are different sizes or colour resolutions
They represent the number of pixels for each different colour represented in the bins

Question 14

Why use colour histograms?

Answer 14

Colour correlates well with class identity
Human vision works hard to preserve colour constancy - presumably because colour is useful
Colour histograms summarise target objects quite well

Question 15

What are some weaknesses for histograms?

Answer 15

Invariant to translation and rotation
Change slowly as viewing direction changes
Change slowly with object size
Change slowly with occlusion

Question 16

What are some problems with colour histograms?

Answer 16

Colour quantisation - noise and/or different camera responses can give similar images, but very different histograms
Histogram resolution - may need many bins to accurately store colour distributions
The illumination may be coloured - same object may generate a different histogram under different lighting
Ignore spatial information

Question 17

What is a histogram intersection?

Answer 17

Measures how much of the query image may be present in the target image (and vice-versa)
A bin in the target histogram can have a larger value than the corresponding query bin (and vice-versa)

Question 18

What did Ballard and Swain use for histogram intersection?

Answer 18

Opponent colour axes & matched images under a range of conditions:
Normal conditions
Varying in view
Varying in image resolutions
Occlusion (of bottom 1/3rd and/or side 1/3rd of image)
Varying in bin resolutions
Varying in light intensity

Question 19

What are region/object-based queries, and how do they work?

Answer 19

Divide the image into windows, see how much of a window matches the query image:
Highlight pixels in the image that are similar to those in the query
Look for regions with lots of these pixels

Question 20

What is the histogram ratio?

Answer 20

Used to determine whether a colour is useful for searching with a query image:
Compute the ratio of corresponding Model and Image histogram bins
If the image has many more pixels of a given colour, Rj is small and that colour is not useful
If the model has more, Rj is 1 and that colour is useful.

Question 21

What is backprojection?

Answer 21

The greater the value of Rj (histogram ratio) the more valuable the colour(s) represented by bin j:
Consider each image pixel
If that pixel maps to histogram bin k, replace the pixel value with a grey value = Rk
Regions with high average brightness are likely to contain the target image