2. TARGET IDENTIFICATION IN A BINARY 
IMAGE. 
The detection of legitimate targets can be described as 
object shape recognition problem - the targets are 
round. Segmentation of the image into 0 = black or 1 
= white ( black for parts of the image below a specific 
intensity level and white for all those on or above) has 
often been successfully used to solve these problems in 
machine vision applications (Haralick, 1985). The 
advantages are: compact images, ease of processing, 
and potential for hardware solutions. The 
disadvantages are that in some circumstances this 
method is likely to prove unreliable because of: variable 
surface reflectivity, non-ideal illumination, the 
possibility of occlusion, the variation of target size, and 
many false targets; i.e. good contrast is needed across 
the whole image. Furthermore, distortion and 
deformation of targets by the imaging process and the 
subsequent digital processing can have an influence on 
the measurement accuracy. The targets chosen are 
required to be distinguishable from the background of 
the object under investigation. Consequently, either 
black or white diffusely reflecting targets are commonly 
used. In this case because the background was of a light 
colour, black targets were chosen. 
2.1 Local Image Normalisation. 
The grey levels of the background of the object are 
seldom constant over the whole image, hence, 
segmentation of the image will often give non-ideal 
results. There are many possible solutions to this 
proue such as: building a mathematical model of the 
ackground image, high pass filtering, dividing the 
image into sections so that each subimage is processed 
separately, or performing a Fourier Transform of the 
image, removing the low frequency components and 
doing an inverse Fourier Transform. Each method has 
its own advantages and disadvantages. 
In practice, using prior knowledge of both the targets 
and the structure being measured can assist in the 
choice of detection algorithms. In the case of this 
application there are many equal sized targets and the 
background intensity changes slowly and provides good 
contrast between the dark targets and the light 
background. To reduce the problem of uneven 
illumination, as shown in Figure 2, affecting the binary 
segmentation; a local area of the image is considered 
for normalisation of the background intensity level. 
The targets were found to occupy a range of pixel sizes 
in the image, from 3X3 to 5X5, therefore the 512X512 
image was divided into 32X32 subimages. The mean of 
all the intensity values was calculated for the whole 
image and the subimage. The subimage was then 
normalised by using Equation 1. 
IM[i,j] = IM[i,j] + (i_mean-1 mean) + C (1) 
Where C is a constant, i mean, and 1 mean are the 
mean intensity values for the whole and subimage 
respectively, and IM[i,j] is the subimage array. 
The advantage of this method is to provide reliable 
thresholding in spite of background intensity variations. 
The value of C can be altered to adjust the image 
background to any desired level. Figure 1. shows the 
original image and Figure 2, shows the inverse intensity 
profi of the section of the image at the position of the 
ine at the top of the image. 
  
Figure 1. Original image. 
  
256 T T T 
224 
192 rF e 
160 
96 
64 | j 
32 
0 1 1 À 
0 64 128 192 256 320 384 448 512 
Intensity / bits 
D 
© 
1 
  
  
  
Pixel position / n 
Figure 2. Intensity profile of marked section. 
If the whole image were segmented above the Eros 
background level, then some of the targets would of 
reduced size, or in the worse cases, not visible at all. 
Figure 3, shows the result of using the image 
normalisation, where it can be seen that segmentation 
will give good definition of the targets. 
  
256 T T T 
224 
192 F 5 
160 
128 = 
96 
Intensity / bits 
32 
0 1 i 1 
0 64 128 192 256 320 384 448 512 
  
  
  
Pixel position / n 
Figure 3. Results of local image normalisation. 
Pre SEE 
tt 
a recent der ri IDR