Automatic classification of protein crystal images - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 354

Canny edge detection [ 14 ] is applied on the image. The Canny edge image is used to extract
edge features, Harris corner [ 15 ] features and Hough [ 16 ] line features. Overall, we obtain a
45-dimensional feature vector for the classifier. All the image feature extraction and classii-
er routines are programmed in Matlab. The following section describes the image processing
and feature extraction in more detail.
FIGURE 2 System diagram.
4 Image preprocessing and feature extraction
The distinguishing characteristics of protein crystals are the presence of straight lines and
quadrangular shapes. Therefore, we focus on extracting geometric features of the objects (or
regions) in the image. As in our previous work [ 13 ], we extract blob features from multiple
binary images and edge features from Canny edge image. In addition to these features, we
apply Hough line transform and extract line related features. Details of our proposed image
processing and feature extraction technique is provided in the following section.
4.1 Green Percentile Image Binarization
Image binarization is a technique for separating foreground and background regions in an im-
age. When green light is used as the excitation source for fluorescence based acquisition, the
intensity of the green pixel component is observed to be higher than the red and blue compon-
ents in the crystal regions [ 7 ] . We utilize this feature for our green percentile image binariza-
tion. Let τ p be the threshold for green component intensity such that the number of pixels in
the image with green component below τ p constitute p % of the pixels. For example, if p = 90, τ 90
is the threshold of green intensity such that 90% of the green component pixels will be less
than τ 90 . Image is binarized using the value of τ p and a minimum gray level intensity condition
τ min = 40. All pixels with gray level intensity greater than τ min and having green pixel compon-
ent greater than τ p constitute the foreground region while the remaining pixels constitute the
background region. As the value of p goes higher, the foreground (object) region in the binary
image usually becomes smaller.
We generate binary images using green percentile thresholds for p =95 and p =99 and extract
region features. Figure 3 shows some sample thresholded images using the two methods.
From the original and binary images in Figure 3 , we can observe that a single technique may
not yield good results for all images. For the images in rows (i) and (ii), the binary images with
p = 95 provide beter representation of the crystal objects. However, for image at row (iii), the
binary image obtained using p = 99 provides beter representation of the crystals.