Literature Reviews - Segmentation of Hand Bone for Bone Age Assessment - page 25

assuming that there are distinct range for object and background themselves. The
value of a valley point is set as threshold.
In image processing, when the histogram of an image is mentioned, typically
it refers to histogram of the values of pixel intensity; the graph of the histogram
represents the number of pixels in an image at each intensity value of the pixel in
the image. If say in an 8-bit grayscale image, there will be 2 8 possible values and
it means that the histogram shows the occurrence frequency of each gray level in
the image. In other words, it is a representation of the image statistics based on the
number of the specific intensity's occurrence.
Histogram analysis is a popular method in automated thresholding [ 47 ]. The
postulation is that the information obtained from the physical shape of the his-
togram of the input image signalizes the suitable threshold value in dividing the
input image into meaningful regions [ 48 ]. Conventionally, the intensity bin in the
valley between peaks is chosen as threshold to reduce the segmentation error rate.
Instead of using manual inspections, by only analyzing the shape of the histogram
and compute the intensity bin that represents the valley, the relatively good thresh-
old value can be found [ 49 ].
However, the main drawback of this technique is that it depends too heavily
on the shape of pixel intensity distribution. Besides, it has no consideration on the
pixels location and the pixel surroundings and this leads to the failure in recog-
nizing the semantic of the input image. This method fails when the input image
does not have distinctly separated intensity distribution between the foreground
and background due to overlapping of intensity as mentioned in last sub-section
of Global thresholding. This category of automated threshold selection performs
thresholding in accordance to the intensity histogram's shape properties. Utilizing
basically the histogram's convex hull and curvature, the intervening valley and
peaks are identified [ 47 ].
This concept is based on the facts that regions with uniform intensity will pro-
duce apparent peaks in the histogram. If only the image have distinct peaks on
each objects in the images, then multiple thresholding is always applicable via
histogram-based thresholding. The favorable shapes of the histogram for the pur-
pose of segmentation are tall, narrow and contain deep valleys. This method is less
influenced by the noise but it has drawbacks such as assuming the pixels intensity
range of the object and background has a certain degree of distinction. If the image
has no distinct valley point in the histogram, this method would fail to separate
the object from the background. The main disadvantage of this histogram-based
thresholding method is the difficulties they meet when they have to identify the
important peaks or valleys in the image used for segmentation and classification.
Selecting the threshold value according to the result of clustering analysis is
one of the major automatic thresholding methods. Depending on the type of clus-
tering analysis, various variations of this category of threshold selection are pro-
posed. One of the most widely implemented methods is Otsu method [ 50 , 51 ]. The
author suggested that the optimal threshold is the intensity point that able to divide
the input image into two groups of intensity when the intra-class variance is mini-
mum or the inter-class variance is maximum [ 52 ].