Literature Reviews - Segmentation of Hand Bone for Bone Age Assessment - page 29

pixels within each cell to determine whether to link the corresponding line seg-
ments in image space into straight line. The central concept is to represent a shape
using the shape's parameter as features instead of image pixels in image [ 63 ].
Hough transform have been widely applied in edge-based segmentation as link-
ing algorithm attributable to its two special capabilities over other approaches:
First, the tolerance towards incompleteness along the boundary line due to occlu-
sion and shape variations; second, the parallel processing capability which is
prominent in real-time application [ 64 ]. The main idea of Hough transform to
determine if a group of pixels are of certain specified shape using shape param-
eters instead of using pixels connectivity contributes to the capability to certain
level of tolerance towards noises [ 65 ]. Besides, the mechanisms of Hough trans-
form to combine independent evidence by analyzing the spatial similarity and
the intensity of parameter peaks contributes to its ability of recognizing partially
deformed shapes. The second capability is due to the nature of Hough transform
in treating each pixel in image space independently [ 66 ]. This makes the parallel
processing by using multiple processing units become possible, implying that real-
time application can be realized. Another added advantage of Hough transform is
its capability to detect several specified shape simultaneously by accumulating evi-
dence produced by each shape in terms of distinct cluster or peak found in the
accumulator array [ 67 ].
Despite the discussed advantages, the classical Hough transform is by no
means a generalized or an optimized shape detector due to several limitations. The
disadvantage of the classical Hough transform is that it is constrained by certain
class of analytically defined shape [ 68 ]. Besides, the input image for the trans-
form is limited to only binary images after being processed by edge detectors.
Also, the resultant connected lines and shape are limited to skeleton lines with
one-pixel thickness. Lastly, the principal limitation of Hough transform is that it
consumes large space storage and requires high computational resources; the com-
putational cost increases exponentially with the parameters dimensionality. In next
paragraph, various improvements have been devised to tackle the aforementioned
limitation.
Numerous improvements to Hough transform through generalization of vari-
ous properties have been proposed over years since its first appearance to address
the aforementioned limitations [ 69 - 73 ]. In 1971, Duda [ 60 ] simplified the compu-
tational complexity of Hough transform by using parameters of radius and angle
instead of slope and intercept and further extends Hough transform into more
general analytically defined curved shape such as circle. In 1981, The Hough
transform was then further generalized using principle of template matching for
applications in which the features of the shapes are not possible to be described
using simple analytic mathematical equations [ 74 ]. Shapiro and Stockman [ 36 ]
proposed the generalization of Hough transform for multiple level images instead
of binary images. Hough transform is then generalized to detect think lines [ 75 ].
The traditional edge-based segmentation techniques have been reviewed. Edge-
based segmentation assumes that the object in image has distinct boundaries. This
assumption is not true in the context of hand bone due to the smooth transition,