The local binary patterns method is a method used for image feature classification in computer vision. In September 1994, we first proposed [43] [44] by T. Ojala, M. pietik äinen, and D. Harwood for texture feature extraction. Later, the combination of the LDA method and the hog feature classifier improved the detection performance on some datasets [45.
The following describes how to extract a feature vector from a local string:
First, the detection window is divided into 16 × 16 small areas (cells). For one pixel in each cell, set the eight points in the ring area (or multiple points in the ring area, 3 points 4. compare the values of the three neighboring regions using the HSV algorithm) clockwise or counterclockwise. If the value of the center pixel is greater than that of the adjacent vertex, the adjacent vertex is assigned 1; otherwise, the value is 0, in this way, each vertex will get an 8-bit binary number (usually converted to a decimal number ). Then calculate the histogram of each cell, that is, the occurrence frequency of each number (assumed to be a decimal number) (that is, a binary sequence about whether each pixel is larger than a neighboring point ), then normalize the histogram. Finally, we connect the statistical histograms of each cell to obtain the texture feature of the whole image. Then we can use SVM or other machine learning algorithms to classify the image.
From: http://blog.csdn.net/abcjennifer/article/details/7429783
Texture classification is a very old topic, but some texture classification methods lay the foundation for future image classification.
First, define the texture image. It is a function that varies with the following variables: texture surface material, reflectivity, illumination, camera and other angles. Currently, texture classification is popular in two ways: global features, such as HSV and Gabor, and local features, such as Harris-Laplace, the local feature-based approach is mainly based on the texton framework, that is, the bag-of-words framework for image classification. Today, we will first introduce Global-based features. Global features are the most effective and simple. Next we will introduce the global features and their variants.
1. HSV
The P-point neighbor with R as the radius, GC as the center, GP as the neighborhood, and the difference between the neighborhood and the center brightness is large or small.
: Change P and R to form a multi-scale HSV
2. Uniform HSV P * (P-1) + 2
In binary encoding, the number of 0-1 conversions is less than or equal to 2;U <= 2:
P = 8, 7*8 + 2 = 58 encoding values, and the rest of U> 2 are classified as a bin
3. The number of images with unchanged rotation is 36.
Since the encoding start point is certain, each binary encoding mode will produce different encoding results after rotation (Cyclic Displacement.
To form a rotation-unchanged encoding mode, we encode the encoded results generated after rotation in the same encoding mode into the same value, that is, the minimum values in these rotation results.
The following code modes are available:
4. Rotate the unchanged uniform HSV P + 1
That is to say, in the uniform sp_in, each row in the first seven rows is rotated and encoded as an encoding value, corresponding to the 1-7 pattern of the first row. In addition, full 1 and full 0A total of 9
-------------------------------- The above is the classic-style Message Service (SLB). The following describes some variants ----------------
1. Increase amplitude information to improve Noise robustness
1. Ltp
Set threshold for binarization
Three-value encoding: normalize the neighborhood of the relative center value in the T range to 0; Quantify the ratio of IC greater than t to 1; Quantify the ratio of IC less than t to-1
Finally, the three-value encoding is converted to positive and negative, and two 8-bit encodings are used as feature vectors;
2. CLBP
The pixel value difference is divided into two considerations: Symbol and amplitude. the encoding of the symbol clbp_s is the same as that of the sp_in (8 bits)
Bitwise amplitude MP encoding (8 bits ):C is the mean of all the MP values in the full image.
GC Code for the central image value (2 bits): Ci indicates the mean value of the full image number.
Finally, we construct a 3D joint histogram clbp_s/M/C, and turn the column into a feature vector.
2. Add local variance information (local contrast)
1.
Obtain the quantitative threshold of local variance in the training set, quantify the local variance, and combine the histogram with the calculation.
Disadvantage: due to different training and test image imaging conditions, the training quantization threshold may not be suitable for the test image.
2.LBP-V
Use the variance of each vertex as the weight of the encoded value to accumulate the histogram (similar to the cumulative gradient amplitude in the direction of sitf ). Principle: The variance is large, and the corresponding region has a large variation, which is a high-frequency region and contributes a lot to the differentiation. Therefore, the corresponding Encoding Weight is large.
3. Add local gradient information (similar to sift)
1. CS-LBP
Encodes the Brightness Difference of the center symmetric point, that is, the graded symbol of the four directions to shorten the encoding length.
2. TP-LBP
Encode the similarity of adjacent patches of a center image to extract patch-based information. The matching center patch and the neighbor patch size of pixels-based information are w * w; the radius of the neighborhood R and number of neighborhood patches are S. The distance A, D (A, B) of Similarity Information is the similarity of A and B patches, and the degree of variation of local patches is encoded.
3. Poem
Encode the gradient information in the patch in the local area.
(1) calculation gradient: Direction and size, M discretization of direction
(2) For each vertex, according to the discrete direction, the cumulative radius is the gradient amplitude (Gaussian weighted) in the r neighbor, and m cumulative gradient amplitude images are formed.
(3) For each image, calculate the r radius and the adjacent P of the HSV to Form M
4. LDP
Changes in the response strength of each vertex's Edge
(1) Calculate the eight-direction edge response
(2) Take the value of MK (the edge response amplitude of K) as the threshold value for binarization to form encoding. C_8_k encoding values
4. Improvement of U-LBP
1. H-LBP
Stacked multi-scale
ULBP classifies all the encoding pairs of U> 2 into a bin and loses the differentiation information.
The larger the radius, the higher the occurrence frequency of Non-uniform encoding, and the larger radius is not uniform encoding. When the smaller radius, it may be in the form of uniform. In this case, convert the encoding to the uniform form in the small radius, until the radius is reduced to the specified size
From: http://blog.sina.com.cn/s/blog_631a4cc401013pkn.html