Image Processing Reference
In-Depth Information
2.1.2 90th Percentile green intensity threshold (g90)
When green light is used as the excitation source for fluorescence based acquisition, the in-
tensity of the green pixel component is observed to be higher than the red and blue compon-
ents in the crystal regions [ 6 ] . 90th percentile green intensity threshold utilizes this feature for
image binarization. First, the threshold intensity ( τ g90 ) is computed as the 90th percentile in-
tensity of the green component in all pixels. This means that the number of pixels in the image
with the green component intensity below this intensity constitutes around 90% of the pixels.
Also, a minimum gray level intensity condition ( t min = 40) is applied. All pixels with gray level
intensity greater than t min and having green pixel component greater than ( τ g90 ) constitute the
foreground region while the rest constitute the background region [ 3 ] . Some sample binary
images of g90 are shown in Figure 1 (g-i).
2.1.3 Maximum green intensity threshold (g100)
This technique is similar to the 90th percentile green intensity threshold described earlier. In
this method, the maximum intensity of green component ( τ g100 ) is used as the threshold intens-
ity for green component. All pixels with gray level intensity greater than t min and having green
pixel component equal to ( τ g100 ) constitute the foreground region. The foreground (object) re-
gion in the binary image from this method is usually smaller than the foreground region from
the other two techniques [ 3 ]. Figure 1 (j-l) shows some sample binary images of g100.
3 DT-Binarize: Selection of best binarization method
using decision tree
In this section, first we describe DT-Binarize that can be used in any image binarization prob-
lem. Then, we briefly define the methods used at intermediate stages of our algorithm. Finally,
we provide application of this method to the protein image binarization problem.
3.1 Overview
Since image binarization is a challenging problem, it is not practical to determine the optimal
threshold value for all cases. There are some weaknesses and strengths of the all image binar-
ization methods [ 9 ] . Based on this fact, in this research, we target an algorithm that selects the
best binarization method rather than a single threshold value. Our goal is to exploit the power-
ful features of different binarization methods and use them whenever they perform well. For
this reason, we propose using a supervised classification method (decision trees) to determ-
ine the best binarization method for any image dataset based on some basic features such as
standard deviation, mean, max intensity, etc.
We first build a training set that is labeled with the best thresholding technique by the ex-
perts. In other words, the best thresholding technique for each image is used as the class label
in that stage. We benefit from ground truth data to select the best binarization method as a
class label. According to our algorithm we first generate a training dataset, in which training
samples are labeled with the binarization methods that provides best binary image. Then in
the training stage, we build the decision tree based on the basic features of the images in the
training dataset. Once we have the decision tree, we are able to determine the best binarization
method for any test image by using the same statistical features. Figure 2 shows the mechan-
ism of the DT-Binarize algorithm. Following steps provide a brief summary of our algorithm:
Search WWH ::

Custom Search