A summary of texture features

Because the classification of the molecular typing effect is not good, so consider extracting features, and think of the teacher asked the last to give him a toolbox, so here incidentally on the medical radiology characteristics of a simple summary

They are used to describe the texture features of the image (although this description is awkward for me, the performance of the classification performance is not good, of course, may be I open the wrong way (; ′⌒ ')

But the reason it's still in use is .... It has an explanatory form in medicine ... Bah

First, they constructed many different matrices based on ROI region I to describe the pixel distribution/correlation

1) Gray-scale co-occurrence matrix (GLCM)

The gray-scale co-occurrence matrix is determined by the position relation of two pixel values in the picture under different angles.

As shown in the following figure, the left I is the original image, we choose the horizontal angle (0 degrees) to construct the gray-scale co-occurrence matrix, Glcm[i,j] is the pixel point of distance I, their pixel difference in I is the number of j-1

We can then compute a series of features based on the GLCM matrix, as shown below

We give a symbolic explanation of some of the features (in view of the poor classification effectiveness, even with the feeling that this interpretation is particularly ... Forced... (T_T)

Variance: When the image gray level changes greatly, the variance performance is bigger

Homogeneity: Image local gray level measurement, if the local uniformity, the value is greater

Contrast: The measurement of local gray level changes, large changes in the image performance is clear, and the value of the larger performance

Dissimilarity: Similar to what contrast expresses, except that this is linearly increasing

Entropy: Characterization of texture complexity, if the larger the texture more complex

......

2 gray-scale travel matrix (GLRLM, I do not make this Chinese translation right, if you do not hope to inform)

Gray travel Matrix used to describe the distribution of pixel values

The image on the left is the original I, the right is a 0-degree structure of the gray-scale co-occurrence matrix, Glrlm[i,j] is the number of pixels in the direction of 0 degrees of continuous J I in the original I in the occurrence of

We can extract the following features

GLD: Measure whether the grayscale is even, the greater the more evenly

RLD: Measure whether the gray stroke is even, the smaller the more evenly

......

In short, the sensory texture feature is to construct a matrix to express the pixel distribution/relevance, and then to describe it by statistical value

Of course, the above is the statistical texture features, in addition to these can also use LBP operator to describe the texture characteristics of the image (when searching for information and found this thing ...) No, I didn't.

In view of the fact that we have not used it, we directly attach the Rachel-zhang to the definition of the right, if there is infringement hope to inform

First, the detection window is divided into 16x16 small area (cell), for each cell in a pixel, the ring adjacent to the 8 points (also can be a circular neighborhood multiple points, such as Figure 3‑4. A clockwise or counterclockwise comparison is performed using the three neighborhood examples shown in the LBP algorithm, if the center pixel value is larger than the adjacent point, the neighbor is assigned a value of 1, otherwise the assignment is 0, so that each point gets a 8-bit binary number (usually converted to a decimal number). The histogram of each cell is then computed, that is, the frequency of each number (assumed to be a decimal number) (i.e. a binary sequence in which each pixel is a bit larger in a neighboring domain), and then normalized to the histogram. Finally, the statistical histogram of each cell is connected, and the LBP texture feature of the whole image is obtained, then the SVM or other algorithms can be used to classify the images.

I looked at the definition and it felt pretty high in the dimensions, a 16x16 cell will work out a 256-dimensional feature, a bit like hog, but it's not likely to be used in its own experiments, because it's a bad way to explain it even if it has good classification effectiveness = = Hey (I'm sad to give up your dream of giving up love is shattered ~ Endure sorrow ~ (; ′⌒ ')

Resources:

http://blog.csdn.net/abcjennifer/article/details/7425483

Https://wenku.baidu.com/view/b1640044561252d380eb6ebe.html