A quiz on image representation and processing in software

Adrian F. Clark

1. The correlation coefficient between two images tells us

   • whether there are gain or offset changes between them

   • how similar they are, ignoring gain or offset changes

   • whether or not two images are identical
   • the amount of noise in the images

   Correlation is a measure which is independent of the means of signals and is normalized by the variances of the two signals, so it is also independent of their relative scales -- it measures only whether the variations in two images relative to their means are in step. The correlation coefficient lies in the range -1 -- +1 and the value describes the similarity of the two images.

2. OpenCV was initially developed in

   • UK

   • Russia

   • Australia
   • USA

   Perhaps surprisingly to some, the early versions of OpenCV were developed in Russia, at Intel's research center in Nizhny Novgorod.

3. Which of the following popular image processing packages provides a menu-based interface?

   • Matlab
   • OpenCV
   • EVE

   • ImageJ

   OpenCV and EVE are libraries to be used by a programmer, while Matlab has a GUI into which one typically types `commands'. However, ImageJ provides menus from which the image processing and analysis operations are chosen.

4. Why would you contrast-stretch an image?

   • to improve its exposure
   • to equalize the image's histogram

   • to spread out the grey levels which are occupied

   • to improve subsequent processing

   The real purpose of contrast-stretching an image is to improve its appearance when seen by a human, with a wider range of grey levels being occupied. This doesn't actually improve its exposure because if only 10 grey levels were occupied before stretching, only ten would be afterwards too, it's just that they are further apart. There are a few cases in which contrast-stretching might help processing (morphology springs to mind) but that's rarely the case. Contrast stretching definitely does not equalize the histogram!

5. How are the colours of an image normally represented in OpenCV?

   • blue, green, red

   • red, green, blue
   • hue, saturation, value
   • blue, red, green

   For reasons known only to its authors, colour images in OpenCV ignore the red, green, blue convention used everywhere else and normally represent images as blue, green, red. Crazy!

6. If using histograms to perform content-based image retrieval, which of the following is suitable for determining similarity?

   • the cross-correlation

   • the square
   • the difference
   • the product

   One can use the sum-squared difference but not the difference, product or square. The cross-correlation, however, does determine similarity.

7. Why does EVE use float values for pixels?

   • it reduces the number of places that the effects of rounding have to be considered

   • that's the way images are stored in disk files
   • it supports a wider range of values
   • it should execute faster

   The use of float values in EVE is largely to reduce the effects of rounding: there are fewer places in a typical piece of code where the effects of rounding need to be considered. There are other advantages too, such as representing a widrer range of values but the rounding one is the most important.

8. What range of values can be stored in an unsigned byte?

   • 0 -- 256
   • -127 -- +127
   • -128 -- +127

   • 0 -- 255

   An unsigned byte represents 256 distinct values, namely 0, 1,... 255.

9. If a histogram has most of its values bunched together at its rght, how does the image appear?

   • well-exposed

   • over-exposed

   • equally-exposed
   • under-exposed

   If the majority of pixels are are the right of a histogram, they have high values and so the image is light and that normally means it is over-exposed.

10. When using OpenCV, how do you index a single pixel?

    • image[y][x][c]

    • image[y,x,c]

    • image[y,x][c]
    • image[c][y,x]

    The correct way to index the image is row-column-channel. The first two of these correspond to the way you'd index a matrix, so this is easy to remember. Then you just have to get the syntax right. There is nothing to distinguish a channel from a pixel location as far as numpy is concerned, and you are giving multiple indices into a single data structure -- the answer image[y][x][c] would be one index into three nested data structures. Figure 3.4 in the lecture notes summarizes the representation.