Literature Reviews - Segmentation of Hand Bone for Bone Age Assessment - page 22

The dynamic threshold, T ( x , y ) is computed via Weighted-Sum operation of the
mean and standard deviation and is denoted as follows, where c is a constant:

(2.5)

T ( x , y ) = m ( x , y ) + c ∗ s ( x , y )

After T is computed, the image undergoes thresholding operation in each pixel.
In the case of binarization in 8 bits input image, the segmentation can be denoted
as follows:

f ( x , y ) < T ( x , y )
f ( x , y ) ≥ T ( x , y )

0
255

IF
IF

f ( x , y ) =

(2.6)

Generally, dynamic thresholding performs better than Global thresholding and
local thresholding. However, it has similar drawback to local thresholding such
that the kernel size is determined manually; the constant in Eq. ( 2.5 ) has to be
selected manually depending on application. Only suitable selection of kernel size
and constant can produce optimum result of segmentation. In addition, dynamic
thresholding consumes much more computational resources relative to local
thresholding and Global thresholding due to its pixel-wise nature. Besides, in per-
forming the neighborhood operations for dynamic thresholding, the padding prob-
lem arises when the kernel approaches the image borders where one or more rows
or columns of the kernel are placed out of the input image coordinates.
In next sub-section, the automated threshold value setting techniques which can
be applied in both Global thresholding and local thresholding are explored and
studied. The implementations of multiple thresholding and local thresholding in
hand bone segmentation is illustrated in next sub-section using automated thresh-
old values selection to demonstrate that the sole implementation of these tech-
niques fail to provide good segmented hand bone.

2.3.4 Automated Thresholding

The main technical issue being frequently discussed is the threshold value selec-
tion: the decision to determine the threshold value in which the object and the
background could be separated as accurate as possible or the decision to select
the threshold value so that the object and the background misclassification rate are
lowest. The result of Thresholding segmentation process depends heavily on this
value. An inaccurate or inappropriate setting of this value will produce disastrous
result in Thresholding segmentation.
For the choice of threshold value, basically, there are two main methods: the
manual threshold selection and the automated threshold selection. Manually deter-
mined threshold value heavily relies on human visual system. Threshold value is
selected using visual perception to partition the object from the background; the
main drawback of this threshold selection is that it involves human subjective