GLCM feature analysis
Two adjacent pixels a certain interval length, or between them having the same gray level , or having different gray levels , if so to identify the statistical distribution of the form of the joint of two pixels , for the image texture analysis It makes sense.
GLCM (GLDM) statistical methods in the early 1970s proposed by R.Haralick and others, it is in the space between each pixel in the image contains the assumed distribution relationship under the premise of image texture information, have proposed breadth of texture analysis methods.
GLCM is defined as each of the gradation from the image pixel i starting point, away from a fixed position (separated by a distance d, orientation of 0 degrees, 45 degrees, 90 degrees, etc.), just gradation value j is the probability that all estimated values can be expressed as a form of a matrix, is referred to in this gray level co.
Here:
- 1. Starting from which gray level i, manually specified.
- 2. spaced distance and direction is also manually specified, the direction may be specified by [0 d].
- 0: [0 d]
- 45 degrees: [-dd]
- 90: [-d 0]
- 135: [-d -d]
- 3. Another artificial gray level j is also specified.
For a slow change in the texture image, its value on the larger diagonal of GLCM; for rapid changes in the texture image, the smaller its value GLCM diagonal, the diagonal sides greater value.
Due to the large amount of data GLCM generally not used directly as a distinguishing characteristic texture, but on some statistic that is constructed as a texture classification characteristic.
Haralick it has been proposed 14 kinds of statistics calculated based on the gray level co: namely: energy, entropy, contrast, homogeneity, correlation, the variance, and the mean, and variance, entropy and difference variance, mean difference, difference entropy , relevant information measure and the maximum correlation coefficient.
Below: https://www.cnblogs.com/8335IT/p/5648445.html
The following diagram shows how the gray level co solving:
The following figure depicts the total eight grayscale
A starting point to a gray level, the gray level is also the target point 1 as an example, the search direction set horizontally spaced 1 (the direction is the horizontal direction, including left and right), GLCM (1,1) pixel value of 1 indicates that only one pair horizontally adjacent.
A starting point to a gray level, the gray level is also the target point 2 as an example, the search direction is provided for the horizontal (the direction is the horizontal direction, including left and right), GLCM (1,2) pixel value of 1 indicates that only two pairs of spaced 1 horizontally adjacent.
For example several commonly used statistics
1. Angle second moment (Angular Second Moment, ASM)
Also known as energy angular second moment, the image gray distribution uniformity of thickness and texture of a metric, an image in which gray scale uniformity of thickness distribution and texture. When uniform texture rule, the larger energy value; conversely element values GLCM similar, smaller energy value.
2.熵（Entropy, ENT)
Entropy measures the amount of information contained random image, showing the complexity of the image. When all the co-occurrence matrix values are equal values or the pixel showing the maximum randomness of the maximum entropy.
3. Contrast
The groove depth of the reaction contrast and sharpness of image texture. The clearer the greater the contrast textures greater contrast.
4. Contrast partial matrix (Inverse Differential Moment, IDM)
Also known as inverse variance matrix inverse difference reflects the degree of clarity and rule texture, texture clear, strong regularity, ease of description, the value larger.
The energy
power conversion reflects the image gray thickness uniformity of distribution and texture. If the gray level co-matrix element values similar to the energy-less, represents a delicate texture; if some large value, other values and small, the energy value is larger. Large energy values indicate a more uniform and regular changes of texture pattern.
6. The inverse variance
inverse variance reflect local variations in texture size, texture if different regions more uniform, slow changes, the inverse of variance will be larger, smaller and vice versa.
7. Correlation
is a measure of the degree of similarity on the gray level image row or column direction, and therefore the size of the reactor is worth local gray correlation, the greater the value, the greater the correlation.
Source: https://blog.csdn.net/guanyuqiu/article/details/53117507