DT-Binarize - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 153

4 Experiments and results
This section provides objective evaluation of each binarization technique using the ground
truth (reference) image dataset that is manually generated by our research group. The correct-
ness of a binary image is calculated using several well-known performance measures. DT-Bin-
arize technique is also compared with the given methods.
4.1 Dataset
The protein crystal images may be grouped into ive main categories: “posetes and spher-
ulites” “needles,” “2D plates,” “small 3D crystals,” and “large 3D crystals.” Distinctive fea-
tures of these categories may be identified as high intense regions, straight edges, and proper
geometric shapes. Our dataset consists of totally 114 protein crystal images that consist of
three subcategories of crystals: “2D plates (40%),” “small 3D crystals (10%),” and “large 3D
crystals (50%).” The size of each image is 320 × 240, and all images have been captured by a
special imaging system under green light. In the following sections, we are going to explain
protein crystal subcategories briefly.
4.1.1 2D Plates
2D Plates have quadrangular shapes. In some specific cases, we may not be able to observe all
the edges of a quadrangular shape because of focusing issues. 2D Plates may have small for
large sizes, and they may be located as a stack of regions. The mean intensities of 2D Plates are
lower than the mean intensities of 3D crystals. This means intensity change between the fore-
ground and the background may not be as significant as for 3D crystals. Figure 5 (a-c) shows
a group of sample images for this category.