Basic image processing operations - Feature Extraction and Image Processing - page 76

normalisation is a linear process and we can return to the original image, should we need
to, or separate pictures, if required.

3.3.4 Thresholding

The last point operator of major interest is called thresholding . This operator selects pixels
which have a particular value, or are within a specified range. It can be used to find objects
within a picture if their brightness level (or range) is known. This implies that the object's
brightness must be known as well. There are two main forms: uniform and adaptive
thresholding. In uniform thresholding , pixels above a specified level are set to white, those
below the specified level are set to black. Given the original eye image, Figure 3.7 shows
a thresholded image where all pixels above 160 brightness levels are set to white, and those
below 160 brightness levels are set to black. By this process, the parts pertaining to the
facial skin are separated from the background; the cheeks, forehead and other bright areas
are separated from the hair and eyes. This can therefore provide a way of isolating points
of interest.

Figure 3.7 Thresholding the eye image

Uniform thresholding clearly requires knowledge of the grey level, or the target features
might not be selected in the thresholding process. If the level is not known, then histogram
equalisation or intensity normalisation can be used, but with the restrictions on performance
stated earlier. This is, of course, a problem of image interpretation. These problems can
only be solved by simple approaches, such as thresholding, for very special cases. In
general, it is often prudent to investigate the more sophisticated techniques of feature
selection and extraction, to be covered later. Prior to that, we shall investigate group
operators, which are a natural counterpart to point operators.
There are more advanced techniques, known as optimal thresholding . These usually
seek to select a value for the threshold that separates an object from its background. This
suggests that the object has a different range of intensities to the background, in order that
an appropriate threshold can be chosen, as illustrated in Figure 3.8 . Otsu's method (Otsu,
1979) is one of the most popular techniques of optimal thresholding; there have been
surveys (Sahoo, 1988; Lee 1990; Glasbey, 1993) which compare the performance different
methods can achieve. Essentially, Otsu's technique maximises the likelihood that the threshold