Images, sampling and frequency domain processing - Feature Extraction and Image Processing - page 34

consider processing its luminance only, often computed in a standard way. In any case, the
amount of memory used is always related to the image size.
Choosing an appropriate value for the image size, N , is far more complicated. We want
N to be sufficiently large to resolve the required level of spatial detail in the image. If N is
too small , the image will be coarsely quantised: lines will appear to be very 'blocky' and
some of the detail will be lost . Larger values of N give more detail , but need more storage
space and the images will take longer to process, since there are more pixels. For example,
with reference to the image of the walking subject in Figure 2.1 (a), Figure 2.2 shows the
effect of taking the image at different resolutions. Figure 2.2 (a) is a 64 × 64 image, that
shows only the broad structure. It is impossible to see any detail in the subject's face.
Figure 2.2 (b) is a 128 × 128 image, which is starting to show more of the detail, but it
would be hard to determine the subject's identity. The original image, repeated in Figure
2.2 (c), is a 256 × 256 image which shows a much greater level of detail, and the subject can
be recognised from the image. (These images actually come from a research programme
aimed to use computer vision techniques to recognise people by their gait; face recognition
would be of little potential for the low resolution image which is often the sort of image
that security cameras provide.) If the image was a pure photographic image, some of the
much finer detail like the hair would show up in much greater detail. This is because the
grains in film are very much smaller than the pixels in a computer image. Note that the
images in Figure 2.2 have been scaled to be the same size. As such, the pixels in Figure
2.2 (a) are much larger than in Figure 2.2 (c) which emphasises its blocky structure. The
most common choices are for 256 × 256 or 512 × 512 images. These require 64 and 256
Kbytes of storage, respectively. If we take a sequence of, say, 20 images for motion
analysis, we will need more than 1 Mbyte to store the 20 256 × 256 images, and more than
5 Mbytes if the images were 512 × 512. Even though memory continues to become
cheaper, this can still impose high cost. But it is not just cost which motivates an investigation
of the appropriate image size, the appropriate value for N . The main question is: are there
theoretical guidelines for choosing it? The short answer is 'yes'; the long answer is to look
at digital signal processing theory.

(a) 64 × 64 (b) 128 × 128 (c) 256 × 256

Figure 2.2 Effects of differing image resolution

The choice of sampling frequency is dictated by the sampling criterion . Presenting the
sampling criterion requires understanding how we interpret signals in the frequency domain .