paper.txt

paper of active shape model notebad

4.1 Implementation of active shape model
4.1.1 Introduction
The active shape model technique was chosen for the prototyping work on this
project. It is a powerful shape description technique. It can represent the typical
shapes and variability of the fluke in the training set. Automatically locating the
fluke in the image is an ill posed problem and so any domain information available
should be captured to help the process. Hence a Point Distribution Model was
implemented.
To build a Point Distribution Model, points were placed around the boundary as
described by Cootes et al. (1995). Bookstein (1991) labeled significant points by
describing their meaning. Hence he called them landmark points. In the case of
whale identification, the fluke was divided into four parts and the first and last
point of each part labeled as major landmark points (see Figure 4.1). These points
mark easily identifiable features of the fluke and have an application-dependent
significance.
19
20 Chapter 4 Chosen approach to automatic fluke location
4.1.2 Placing model points
To help build the model, a prototype system was constructed in Matlab in which
the user can enter as many points as necessary along the boundary for each part
of the fluke by using the left mouse button. The right mouse button may be
used to specify the endpoint of each part. The five major landmark points can be
observed in Figure 4.1, where they are represented by green, round circles. Placing
the points manually is very time consuming. Note however that the points are
only placed manually during the training phase when building a point distribution
model and would not be required later. For capturing the shape variability reliably,
it is important that the labeling is done correctly along the boundary on each
training shape.
4.1.3 Sampling
To ensure that every image is described by the same number of points, the system
can equally space the rest of the points along the connecting boundaries. The
length of each boundary part is calculated and divided by the number of points to
sample. The location for each point is calculated. When entering the points and
when sampling, a line is drawn around the fluke and the new points are represented
by red x抯 for easy observation as shown in Figure 4.1. Due to the division of the
fluke into parts, major landmark points remain unchanged.
Figure 4.1: 104 landmark points on the boundary of the fluke with the green
circles representing the major landmark points.
An early approach required entry of six points per part, but then the system was
Chapter 4 Chosen approach to automatic fluke location 21
improved to allow the user to enter as many points as necessary in order to model
distinctive shapes without loosing important information. The system allows for
the number of sampling points to be adjusted. Each structure is now represented
by 104 points, which seems to describe the fluke with sufficient accuracy.
4.1.4 Aligning the training shapes
Cootes et al. (1995) compare equivalent points from different shapes by scaling,
rotating and translating the training shapes so that they correspond as closely as
possible. As a computational compromise for aligning the training shapes, the
fluke notch was set to zero in all training shapes as a new origin. Therefore all
training shapes are translated to the same origin, without computing scale and
rotation. The scale and rotation could be included later on if required.