Image Processing Reference
In-Depth Information
ation method for these types of images in the classification process. Improper binary image
may lose some important information or it may keep some unnecessary information such as
tiny noisy pixels. Furthermore, incorrect thresholding method may lose a blurred crystal in
the image or it may crop a regular object around its borders leading incorrect classification
In our previous work [ 6 ] , we used three thresholding techniques (Otsu's Threshold (Otsu),
90th Percentile Green Intensity Threshold (g90), and Max Green Intensity Threshold (g100))
together not to lose any informative feature for classification of protein crystallization images.
However, this also leads incorporation of unnecessary features that may yield incorrect clas-
siication results. To avoid this problem, in this study, we propose an alternative approach,
DT-Binarize , recently introduced in [ 7 ]. It selects the best thresholding technique for a particu-
lar image using decision tree.
In DT-Binarize, we train a decision tree using some basic features of the protein images on
our pre-labeled samples. Each label indicates a different thresholding technique that properly
its for that particular image. In the test stage, the best thresholding method is selected for a
given test image using the same features. Our technique tries to select the most useful and re-
liable binarization methods for the protein crystallization images. In this way, the complexity
of our system may be reduced since we are dealing with less number of features (i.e., features
from a single thresholded image are used rather than from multiple thresholded images).
This research uses protein crystallization images dataset provided by iXpressGenes, Inc. As
our earlier work, we classify the protein images into three main groups (non-crystals, likely-
leads, and crystals). Each category has its own specific characteristics that need to be con-
sidered independently. In this paper, we focus on “crystals” only and propose a solution to
select the best thresholding technique for each image.
The rest of the chapter is structured as follows. The image binarization techniques are de-
scribed in Section 2 . Our approach to select the best binarization technique is explained in Sec-
tion 3 . Experimental results are provided in Section 4 . Finally, our chapter is concluded in Sec-
tion 5 .
2 Background
2.1 Image Binarization Methods
Image binarization is a technique for separating foreground and background regions in an
image. For the protein images consisting of crystals, the crystal regions are expected to be
represented as the foreground in the binary images. Some of the sample protein images are
provided in Figure 1 (a-c). While a thresholding technique may perform well for an image, it
may not perform as good as other thresholding techniques for another image. Thus, we con-
sider three image thresholding techniques described below: Otsu's threshold (Otsu), 90th per-
centile green intensity threshold (g90), and max green intensity threshold (g100).
Search WWH ::

Custom Search