Image Processing Reference
In-Depth Information
again represents a shape as a set of landmark points and uses a set of training data to
establish the potential range of variation in the shape. One major difference is that AAMs
explicitly include texture and updates model parameters to move landmark points closer to
image points by matching texture in an iterative search process. The essential differences
between ASMs and AAMs include:
1.
that ASMs use texture information local to a point, whereas AAMs use texture information
in a whole region;
2.
that ASMs seek to minimise the distance between model points and the corresponding
image points, whereas AAMs seek to minimise distance between a synthesised model
and a target image;
3.
that AAMs search around the current position - typically along profiles normal to the
boundary, whereas AAMs consider the image only at the current position.
A recent comparison (Cootes, 1999) has shown that although ASMs can be faster in
implementation than AAMs, the AAMs can require fewer landmark points and can converge
to a better result, especially in terms of texture (wherein the AAM was formulated). We
await with interest further developments in these approaches to flexible shape modelling.
6.6
Further reading
The majority of further reading in finding shapes concerns papers, many of which have
already been referenced. An excellent survey of the techniques used for
feature extraction
(including template matching, deformable templates etc.) can be found in Trier (1996)
whilst a broader view was taken later (Jain, 1998). A comprehensive survey of flexible
extractions from
medical
imagery (McInerney, 1996) reinforces the dominance of snakes
in medical image analysis, to which they are particularly suited given a target of smooth
shapes. (An excellent survey of history and progress of medical image analysis has appeared
recently (Duncan, 2000).) Few of the textbooks devote much space to shape extraction and
snakes are too recent a development to be included in many textbooks. One text alone is
dedicated to shape analysis (Otterloo, 1991) and contains many discussions on symmetry.
For implementation, Parker (1994) only includes C code for template matching and for the
HT for lines, but no more. A visit to Dr Cootes' website suggests that a text might be on
the way on flexible shape modelling, so we can await that with interest.
6.7
References
Bamford, P. and Lovell, B., Unsupervised Cell Nucleus Segmentation with Active Contours,
Signal Processing
,
71
, pp. 203-213, 1998
Berger, M. O., Towards Dynamic Adaption of Snake Contours,
Proc. 6th
Int. Conf. on
Image Analysis and Processing
, Como, Italy, pp. 47-54, 1991
Cham, T. J. and Cipolla, R., Symmetry Detection Through Local Skewed Symmetries,
Image and Vision Computing
,
13
(5), pp. 439-450, 1995
Cohen, L. D., NOTE: On Active Contour Models and Balloons,
CVGIP: Image Understanding
,
53
(2), pp. 211-218, 1991
Cohen, I., Cohen, L. D. and Ayache, N., Using Deformable Surfaces to Segment 3D
