2
Electronic Supplementary Material
for
Heta Mattila, Pertti Valli, Tapio Pahikkala, Jukka Teuhola, Olli S. Nevalainen, Esa Tyystjärvi
Comparison of Chlorophyll Fluorescence Curves and Texture Analysis for Automatic Plant Identification
Calculation of Local Binary Patterns (LBPs)
was computed by using the gray value of each center pixel and its eight neighbors as shown in Fig. S1 where the circle is approximated by a 3 x 3 pixel square. First the neighborhood was thresholded by the brightness value of c. Each neighbor pixel was represented by a single bit (0/1 = black/white or darker/lighter than the center). To map each LBP to a single number, the thresholded 0/1-neighbourhood pixel values were multiplied by the corresponding binomial weights (Fig. S1), and the number was obtained by summing the values of the eight pixels. This number is invariant on the gray-scale level. Rotation invariance was finally obtained by taking the minimum LBP number of the 8 alternative rotations when computing the LBP (Pietikäinen et al. 2000).
Fig. S1. Computation ofusing a circularly symmetric neighbor set.
As shown by Ojala et al. (2002), for P = 8 there are 36 unique rotation-invariant LBPs that can be considered as feature instances. The 36 different LBPs were clustered to 2 groups with no 0/1 transitions in the perimeter (all black or all white), 7 groups with 2 transitions and 1 group with 4 or more transitions. The first 9 groups are the uniforms, and the last group contains 27 miscellaneous patterns (Ojala et al. 2002) (Fig. S2). These binary texture features were further smoothed by replacing them with the average feature values of the pixels in the neighborhood of the pixel under consideration.
Fig. S2. The nine unique rotation invariant uniform local binary patterns in the case of.
The operator, where riu2 indicates that the uniforms have maximally 2 transitions, was implemented by using Java programming language harnessed with ImageJ (http://rsbweb.nih.gov/ij/docs/intro.html. Accessed 11 February 2013). For feature extraction, the value was calculated for each pixel in the image and mapped to the corresponding uniform value by using a mapping table. The ImageJ library was used to read and display the RGB images in HSB color space. The brightness channel was used to compute the LBP values. A circular neighborhood with a radius 1 pixel was calculated by extracting a 3 x 3 pixel matrix for every image pixel (Fig. S1). The values were then calculated and mapped to corresponding uniform values.