Automatic classification of protein crystal images - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 357

Table 2
Edge Features

Symbol Description

η Number of graphs (connected edges)

η 1 Number of graphs with a single edge

η 2 Number of graphs with two edges

η c Number of graphs whose edges form a cycle

η p Number of line normals

μ l Average length of edges in all segments

S l Sum of lengths of all edges

l max Maximum length of an edge

4.4 Corner Features
Corner points are considered as one of the uniquely recognizable features in an image. A
corner is the intersection of two edges where the variation between two perpendicular direc-
tions is very high. Harris corner detection [ 15 ] exploits this idea and it basically measures the
change in intensity of a pixel ( x , y ) for a displacement of a search window in all directions. We
apply Harris corner detection and count the number of corners as the image feature.

4.5 Hough Line Features
Hough transform is a very popular technique in computer vision for detecting certain class of
shapes by a voting procedure [ 16 , 18 ] . We apply Hough line transform to detect lines in an im-
age. Once the lines are detected, we extract the following two line features—number of Hough
lines and the average length of line.

5 Experimental results
Our experimental dataset consists of 212 expert labeled images. The images are hand-labeled
by an expert into four different categories: needles, small crystals, large crystals, and other
crystals. The proportion of these classes are 24%, 20%, 35% and 21% respectively. For each im-
age, we apply green percentile binarization with p = 95 and p = 99. From each binary image,
we extract 17 region features. Likewise, we extract 8 edge related features, 1 Harris corner fea-
ture, and 2 Hough line features. Therefore, we extract a total of 2 × 17 + 8 + 1 + 2 = 45 features
per image. On a Windows 7 Intel Core i7 CPU @2.4 GHz system with 12 GB memory, it takes
around 232 s to extract features for 212 images. Thus the time for feature extraction is around
1.1 s per image.
We group the feature sets into three categories—region features, edge/corner/line features,
and combined features. For the classification, we test using decision tree and random forest
classiier. Table 3 shows the classification accuracy using the selected classifiers and feature
sets. The values are computed as the average accuracy over 10 runs of 10-fold cross validation.