Feature extraction by shape matching - Feature Extraction and Image Processing - page 162

a description of a shape, such as its position and size, whereas detection of a shape merely
implies knowledge of its existence within an image.
The techniques presented in this chapter are outlined in the table below. In order to
extract a shape from an image, it is necessary to identify it from the background elements.
This can be done by considering the intensity information or by comparing the pixels
against a given template. In the first approach, if the brightness of the shape is known, then
the pixels that form the shape can be extracted by classifying the pixels according to a fixed
intensity threshold. Alternatively, if the background image is known, then this can be
subtracted to obtain the pixels that define the shape of an object superimposed on the
background. Template matching is a model-based approach in which the shape is extracted
by searching for the best correlation between a known model and the pixels in an image.
There are alternative ways to compute the correlation between the template and the image.
Correlation can be implemented by considering the image or frequency domains. Additionally,
the template can be defined by considering intensity values or a binary shape. The Hough
transform defines an efficient implementation of template matching for binary templates.
This technique is capable of extracting simple shapes such as lines and quadratic forms as
well as arbitrary shapes. In any case, the complexity of the implementation can be reduced
by considering invariant features of the shapes.

Table 5.1 Overview of Chapter 5

Shape extraction method Technique

Pixel brightness Image thresholding

Image subtraction

Template matching Intensity template

Image and Fourier domains Binary Hough Lines

templates transform Quadratic forms Invariance

Arbitrary shapes

5.2 Thresholding and subtraction

Thresholding is a simple shape extraction technique, as illustrated in Section 3.3.4 where
the images could be viewed as the result of trying to separate the eye from the background.
If it can be assumed that the shape to be extracted is defined by its brightness, then
thresholding an image at that brightness level should find the shape. Thresholding is
clearly sensitive to change in illumination: if the image illumination changes then so will
the perceived brightness of the target shape. Unless the threshold level can be arranged to
adapt to the change in brightness level, any thresholding technique will fail. Its attraction
is simplicity : thresholding does not require much computational effort. If the illumination
level changes in a linear fashion, then using histogram equalisation will result in an image
that does not vary. Unfortunately, the result of histogram equalisation is sensitive to noise,
shadows and variant illumination: noise can affect the resulting image quite dramatically
and this will again render a thresholding technique useless.
Thresholding after intensity normalisation (Section 3.3.2) is less sensitive to noise,
since the noise is stretched with the original image, and cannot affect the stretching process
by much. It is, however, still sensitive to shadows and variant illumination. Again, it can