The structure represents texture-based local properties of micro-texture and macro textures represent the spatial texture of narrow properties. These properties are not similar between the image pixels. The statistical features based method builds relationship among the gray levels. One-pixel based classifiers known as first order derivative and more than two pixels based classifier is known as second order derivative. The second order derivative is one of the greatest derivate in texture analysis. In biomedical and remote sensing fields these kinds of textures are evaluated.
The feature of the first order histogram provides different statistical properties such as four statistical moments of the intensity histogram of an image. These depend only on entity pixel values and not on the interaction or else co-occurrence of neighboring pixel values. The first order histogram statistics are mean, skewness, kurtosis and entropy. The GLCM is used to extract the second order texture in sequence from the images.
Mean
The mean describes the center value of the intensity pixels and it is denoted by the image features:Mean=∑b=0l−1p(s).(3.9)
Skewness
Skewness (sw) is a computation of the asymmetry of the probability distribution of a real-valued random variable. The skewness value is negative or positive. The probability is considered as total number of neighborhood (p(s)) centered pixels by number of gray level pixels(s):sw=1σs3Σb=01−1(s−s¯)3 p(s).(3.10)
Kurtosis
The kurtosis is the measurement of “peakedness” of the probability distribution of real value-based random variable:kw=1σs4Σb=01−1(s−s¯)4 p(s)−3.(3.11)
Entropy (En)
The concept of entropy comes from thermodynamics. Entropy is used to calculate the randomness and used to describe the texture of the input image. This value is between maximum of all elements of co-occurrence matrix. The pixels of i and j characterize the coefficient of concurrence matrix J(i,j) and S represented as dimension of co-occurrence matrix. It is defined asEn=∑i=0S−1∑j=0s−1J(i,j)(−ln(J(i,j)).(3.12)
Contrast (Cn)
The contrast is a measure consisting of pixel intensity and its neighbor over the image and it belongs to the brightness of the object:Cn=∑i=0S−1∑j=0s−1(i−j)2J(i,j).(3.13)
Energy (Eg)
Energy is defined as extension of pixel pair replication and calculates the uniformity:Eg=∑i=0S−1∑j=0s−1J2(i,j).(3.14)
After segmentation, all gray level co-occurrence matrix features are extracted from the segmented images and the weed and plant area are distinguished using machine learning techniques. The neural network and support vector machine (SVM) classifier help to detect weed infected area with accuracy. The gray level co-occurrence matrix feature values are given in Table 3.1.
| Segmented image | Mean | Skewness | Kurtosis | Entropy | Contrast | Energy |
|---|---|---|---|---|---|---|
| Image 1 | 0.2224 | 11.1851 | 1.0081e+004 | 6.6791 | 0.3425 | 0.2987 |
| Image 2 | 0.1592 | 8.9473 | 2.0918e+004 | 7.6791 | 0.9402 | 0.1645 |
| Image 3 | 0.2088 | 15.9818 | 1.5981e+004 | 9.6791 | 0.3501 | 0.2077 |
About 60 image features are taken for SVM and neural network classifiers to detect the weed and plants dissimilarities. Both classifiers are applied to detect the weed and plant regions via bounding boxes as shown in Figure 3.4. These kinds of machine learning algorithms are used in robotics for selecting the region and spraying herbicides automatically.
Leave a Reply