Literature Reviews - Segmentation of Hand Bone for Bone Age Assessment - page 37

The advantages of active contour compared with previously discussed methods:
1. Process the image pixels in specific areas only instead of the entire image and
thus enhances the computational efficiency.
2. Impose certain controllable prior information.
3. Impose desired properties, for instance, contour continuity and smoothness.
4. Can be easily governed by user by manipulating the external forces and
constraints.
5. Respond to image scale accordingly with the assistance of filtering process.
Disadvantages of classical active contour model:
1. Not specific enough to be implemented in specific problem as the shape of the
targeted object is often not recognized by the algorithm.
2. Unable to segment multiple objects.
3. High sensitivity to environmental noises in image.
4. High dependency towards intensity gradient along the edges.
5. Do not consider the region information of the targeted objects.
6. High dependency on initial guessed point location. If the initial snake is not
sufficiently close to targeted object boundaries, then points in snake can hardly
attach the boundaries.
7. Difficult to grow into concavity.
8. Do not have a global shape controller that constraints the shape of contour from
deviating from allowable shape of the targeted object.
To sum up, the regularizing terms adopted in active contour model is useful
in stabilizing the contour but the robustness is limited as the imposed constraints
generally tend to smooth and shorten the contour unless stronger external energy
is involved; this scheme is often too general and inadequate. Therefore, a more
specifically designed scheme that capable of incorporating more finely tuned prior
knowledge about the class of targeted object is required.
2.7.2 Active Shape Model (ASM)
ASM is a model founded on statistical theory where the variations of the shape of
the objects can be captured via training procedure using labeled object's contour in
the image in set points representation. Activating the trained contour will deform
the contour fitting the targeted object in the image. Cootes et al. [ 97 ] developed the
model. Generally, it works by searching the best position of initial points that are
surrounding the object, and then updating these positions until the stopping criteria
are achieved through iterations. Ever since the technique is proposed until recently,
it has been extensively applied in various fields such as facial recognition [ 98 - 100 ],
object tracking [ 78 , 101 - 104 ] and medical image processing [ 11 , 105 - 107 ].
The first step in ASM is to establish a shape model by a set of shape examples
from training images. This process is known as the training phase of ASM. The