Feature extraction by shape matching - Feature Extraction and Image Processing - page 161

5

Feature extraction by
shape matching

5.1 Overview

High-level feature extraction concerns finding shapes in computer images. To be able to
recognise faces automatically, for example, one approach is to extract the component
features. This requires extraction of, say, the eyes, the ears and the nose, which are the
major face features. To find them, we can use their shape: the white part of the eyes is
ellipsoidal; the mouth can appear as two lines, as do the eyebrows. Shape extraction
implies finding their position, their orientation and their size. This feature extraction process
can be viewed as similar to the way we perceive the world: many books for babies describe
basic geometric shapes such as triangles, circles and squares. More complex pictures can
be decomposed into a structure of simple shapes. In many applications, analysis can be
guided by the way the shapes are arranged. For the example of face image analysis, we
expect to find the eyes above, and either side of, the nose and we expect to find the mouth
below the nose.
In feature extraction, we generally seek invariance properties so that the extraction
process does not vary according to chosen (or specified) conditions. That is, techniques
should find shapes reliably and robustly whatever the value of any parameter that can
control the appearance of a shape. As a basic invariant , we seek immunity to changes in
the illumination level: we seek to find a shape whether it is light or dark. In principle, as
long as there is contrast between a shape and its background, the shape can be said to exist,
and can then be detected. (Clearly, any computer vision technique will fail in extreme
lighting conditions, you cannot see anything when it is completely dark.) Following
illumination, the next most important parameter is position : we seek to find a shape
wherever it appears. This is usually called position- , location- or translation-invariance .
Then, we often seek to find a shape irrespective of its rotation (assuming that the object or
the camera has an unknown orientation): this is usually called rotation- or orientation-
invariance . Then, we might seek to determine the object at whatever size it appears, which
might be due to physical change, or to how close the object has been placed to the camera.
This requires size- or scale-invariance . These are the main invariance properties we shall
seek from our shape extraction techniques. However, nature (as usual) tends to roll balls
under our feet: there is always noise in images. Also since we are concerned with shapes,
note that there might be more than one in the image. If one is on top of the other, it will
occlude , or hide, the other, so not all the shape of one object will be visible.
But before we can develop image analysis techniques, we need techniques to extract the
shapes. Extraction is more complex than detection , since extraction implies that we have