Image Processing Method for Automatic Discrimination of Hoverfly Species

Figure 2

Illustration of HOG features on the example of one of the images from the training/test set, Section 2.2. In (a)–(c) are visualized HOG features corresponding to different values of cell size: pixels (a), pixels (b), and pixels (c). In each case (a)–(c), for visualization are used two images: one depicting discrete histograms of gradient orientation on the level of single cells in the image (images consisting of small histograms with blue, green, and red bars) and the second image which represents its grayscale counterpart (on the level of each cell, i.e., histogram, are drawn lines which correspond to particular orientation of image gradient, while intensity of lines, i.e., their normalized grayscale value, corresponds to the magnitude of the image gradient in the given direction after gradient discretization). Blue arrow in (a) indicates equivalence between magnified histogram and its grayscale counterpart, with lines associated with the bars in the histogram. Different color of bars in the histograms is used to ease distinction between allowed discrete orientations, 9 values in the range 0–180^°. Green arrows in the upper part of (a) describe the process of image gradient computation using Sobel filter, while the blue grid depicts the cell size, also overlayed over gradient images in (b) and (c).

(a) hog8

(b) hog16

(c) hog32