A texture classification method based on MRLBP
By using a combination of amplitude sorting and multi-scale histograms in the local binary mode, the problem of insufficient utilization of amplitude information and fixed weight sorting in texture classification is solved, and the accuracy and robustness of texture classification are improved. Large-scale or real-time scenarios.
Patent Information
- Application Number
- CN202111361874.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-17
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2041-11-17
AI Technical Summary
The existing local binary modes lack the utilization of amplitude information in texture classification and the use of fixed weights for sorting, resulting in low classification accuracy.
A local binary mode (MRLBP) algorithm based on amplitude sorting is proposed. By sorting symbols with larger amplitude information, the symbols with larger amplitude have higher weights, and the multi-scale histogram and SVM classifier are used for texture. Classification.
It improves the accuracy of texture classification, maintains the rotation invariance and noise robustness of the algorithm, and reduces the feature dimension and calculation complexity, making it suitable for large-scale or real-time classification scenarios.
Smart Images

Figure CN114140643B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of image processing, and in particular relates to a texture classification method based on local binary patterns, and improves the accuracy of texture classification by improving a local binary pattern algorithm. Background Art
[0002] With the informatization of society, more and more application research and development are closely related to the field of image processing. Texture classification, as one of the basic problems in the field of image processing, has also received extensive attention and research. As a local feature description operator, Local Binary Pattern (LBP) is widely used in texture classification and has achieved remarkable results due to its simple principle, easy implementation, low computational complexity, certain illumination invariance, and no need for advance training. Therefore, the research and improvement of local binary patterns has very important practical significance. Since the local binary pattern was proposed in 1994, a large number of variants have appeared. The DLBP algorithm proposed by Liao et al. improved the original local binary pattern uniform pattern from the perspective of the original local binary pattern uniform pattern. The MBP algorithm proposed by Hafiane et al. improved the neighborhood structure and quantization threshold. The LTP algorithm proposed by Tan et al. changed the quantization level. The BRINT algorithm proposed by Liu et al. improved the robustness to noise. Most of the variants that have appeared so far, like the above variants, have been improved from different angles, which have improved the classification accuracy to a certain extent. However, none of them considers the rationality of using fixed weights to sort local binary patterns, and the current variants do not make comprehensive use of amplitude information. For example, the CLBP algorithm proposed by Guo et al. only uses the sign information of the amplitude relative to the center pixel. Summary of the invention
[0003] The purpose of the present disclosure is to provide a texture classification method based on MRLBP, and propose a local binary pattern based on amplitude sorting to solve the problem of insufficient use of amplitude information and sorting using fixed weights when using the original local binary pattern for texture classification, thereby improving the accuracy of texture classification. Specifically, the following steps are included:
[0004] Step 1, read the image in the texture data set, if the image is not a grayscale image, convert it into a grayscale image first;
[0005] Step 2, processing the image by using the amplitude-sorted local binary pattern algorithm, and obtaining an image processed by the amplitude-sorted local binary pattern (MRLBP) algorithm by using the sign information and the amplitude information;
[0006] Step 3, extracting features from the processed image, first pre-training the image data after the amplitude-sorted local binary pattern processing to effectively reduce the feature dimension;
[0007] Step 4, then use a multi-scale histogram to combine histograms of different scales in a connected manner to obtain a higher classification accuracy;
[0008] Step 5: finally, the obtained histogram is normalized to remove the dimension of the feature, so that different features are mapped to the same interval and the convergence speed of the classifier is improved;
[0009] Step 6: Use SVM classifier to classify the texture image.
[0010] Further, the implementation of the amplitude sorting-based local binary pattern algorithm described in step 2 includes the following steps:
[0011] Step 21, selecting a local area of a suitable scale in the texture image currently read;
[0012] Step 22, using bilinear interpolation method, calculate the grayscale values of P neighborhood pixels of the circularly symmetric neighborhood set in the local area, where the coordinates of the neighborhood pixels are expressed as (x+Rcos(2πp / P), y+Rsin(2πp / P)), (x, y) are the coordinates of the center pixel, R is the scale, p is the index value of the current neighborhood pixel, and P is the number of neighborhood pixels;
[0013] Step 23, for the central pixel of each local area, the gray value of the central pixel is subtracted from the gray value of the neighboring pixel to obtain the symbol information and amplitude information of the central pixel; then the symbols are sorted according to the amplitude, that is, different weights are assigned. The larger the amplitude, the higher the weight assigned to the corresponding symbol. The obtained binary pattern is the local binary pattern of the central pixel sorted based on the amplitude. For the specific implementation method, see formulas (8) and (9):
[0014]
[0015]
[0016] in, Formula (8) is the local binary pattern of the central pixel based on the amplitude sorting, and Formula (9) is the function used to obtain the weight corresponding to each symbol in Formula (8), where g p and g i Both refer to the grayscale value of the neighborhood pixels, g c Refers to the gray value of the center pixel;
[0017] Step 24, performing the same operation on each central pixel of the image to obtain an image processed by the local binary pattern based on amplitude sorting.
[0018] Furthermore, the calculation formula of the bilinear interpolation method in step 22 is:
[0019]
[0020]
[0021]
[0022] That is, using Q 11 (x1,y1),Q 12 (x1,y2),Q 21 (x2,y1) and Q 22 The four gray values (x2, y2) are G(Q 11 ), G(Q 12 ), G(Q 21 ) and G(Q 22 ) are interpolated in the x-axis direction to obtain the grayscale values G(R1) and G(R2) of R1(x,y1) and R2(x,y2), and then these two grayscale values are interpolated in the y-axis to obtain the grayscale value G(P) of the required point P(x,y).
[0023] Furthermore, the pre-training in step 3 includes the following steps:
[0024] Step 31, counting the probability of each pattern appearing in the entire texture data set, that is, obtaining a histogram of 256 patterns on the entire training set, and counting the total number of patterns, thereby obtaining the probability of each pattern appearing in the entire training set;
[0025] Step 32, add the probabilities from high to low, and keep the patterns whose probabilities and sums are within K;
[0026] Step 33, for the mode whose probability is not within K, merge it into the mode that is closest to it; the closest means that the difference between the two modes is the smallest. The specific method is to traverse each mode and determine whether the probability corresponding to the current mode is within K. If it is within K, no processing is performed and the traversal continues; if it is not within K, calculate the mode that is closest to the current mode with a probability within K, and merge the current mode into this mode;
[0027] Step 34, processing the image after the amplitude-sorted local binary pattern processing according to the pre-training rules of the corresponding texture data set can reduce the feature dimension to a certain extent.
[0028] Furthermore, the combination of the histograms of different scales includes the following three cases:
[0029] Combination of one-scale histogram, two-scale histogram and three-scale histogram;
[0030] Combination of one-scale histogram, two-scale histogram, three-scale histogram and four-scale histogram;
[0031] The one-scale histogram, the two-scale histogram, the three-scale histogram, the four-scale histogram and the five-scale histogram are combined.
[0032] Furthermore, the P value of the P neighborhood pixels in step 22 is 8, and the angles of these 8 neighborhood pixels on the circularly symmetric neighborhood set are 0°, 45°, 90°, 135°, 180°, 225°, 270° and 315° respectively.
[0033] Furthermore, when weights are assigned to symbols according to the amplitude in step 23, if the amplitudes are the same, the symbols with the same amplitude are assigned the same weights.
[0034] Furthermore, the probability and K include five values: 0.5, 0.6, 0.7, 0.8, 0.9 and 1.0.
[0035] Furthermore, in step 5, the histogram is normalized using the Z-score normalization method.
[0036] Furthermore, in step 6, the SVM classifier kernel function selects the radial basis function RBF, the penalty factor C is 10.0, and the coefficient Gamma is 0.01.
[0037] Compared with the prior art, the beneficial effects and advantages of the present invention are:
[0038] 1. The present invention proposes a local binary pattern based on amplitude sorting, which solves the problem of insufficient utilization of amplitude information and sorting using fixed weights when using the original local binary pattern for texture classification, and achieves a higher texture classification accuracy.
[0039] 2. The present invention abandons the use of fixed weights and uses amplitude to sort symbols. When the image is rotated, the local binary pattern of each central pixel remains unchanged, that is, the algorithm has good rotation invariance.
[0040] 3. The present invention adopts the idea of assigning higher weights to larger amplitudes. Therefore, in general, noise interference will only affect low-weight binary values, and the impact on the entire local binary pattern is minimized, that is, the algorithm has a certain degree of noise robustness.
[0041] 4. The present invention retains the advantage of low computational complexity of the original local binary pattern, has good real-time performance, and is suitable for large-scale classification or scenarios with high real-time requirements. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 It is an overall implementation block diagram of the present invention.
[0043] Figure 2 It is a flow chart of obtaining weights in the local binary pattern algorithm based on amplitude sorting of the present invention.
[0044] Figure 3 It is a specific example of the local binary pattern algorithm based on amplitude sorting of the present invention.
[0045] Figure 4 It is a black and white chessboard diagram used by the present invention to illustrate the fixed weight problem. In the figure, (a) is the original image, (b) is the image after the original local binary pattern processing, and (c) is the image after the local binary pattern processing based on amplitude sorting. DETAILED DESCRIPTION
[0046] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.
[0047] In order to solve the problem of insufficient utilization of amplitude information and fixed weights for sorting when using original local binary patterns for texture classification, the present invention proposes a texture classification method based on MRLBP, which processes images through a local binary pattern algorithm based on amplitude sorting to improve the accuracy of texture classification.
[0048] The specific implementation method and operation process of the present invention include the following 7 steps:
[0049] Step 1, read the image in the texture dataset. In order to compare the algorithms, the seven most commonly used texture datasets are selected, including Outex_TC10, Outex_TC12_000, Outex_TC12_001, CUReT, UIUC, UMD and KTHTIPS2b. The image information of these seven datasets is shown in Table 1. Among them, the CUReT dataset uses the processed version proposed by Varma et al., and 92 images with an observation angle less than 60° and easy to crop into 200×200 are extracted from each class of the original CUReT dataset and converted into grayscale images. For CUReT, UMD and UIUC datasets without pre-defined training and test sets, half of the samples in each class are randomly selected as training sets, and the remaining half are used as test sets. The experiment is repeated 50 times and the average is taken as the final result. For the KTHTIPS2b dataset, the first three samples of the four samples in each class are selected as training sets, and the remaining one sample is used as the test set. All images of the seven datasets are 8-bit grayscale images.
[0050] Table 1
[0051]
[0052]
[0053] Step 2, calculate the information contained in the sign and amplitude. p and the amplitude m p The expression of can refer to formula (1) and (2), where g p is the gray value of the neighborhood pixel, g c is the gray value of the center pixel, and formula (3) is the specific representation of the function in formula (1):
[0054] s p =s(g p -g c ) (1)
[0055] m p =|g p -g c | (2)
[0056]
[0057] The statistics of each symbol s in the data set p and amplitude m p The probability of occurrence is calculated using the information entropy formula:
[0058]
[0059] Where N is the number of all signs or magnitudes, P(xi ) is the probability of each symbol or amplitude occurring.
[0060] The above operations are performed on the five texture datasets Outex_TC10, Outex_TC12_000, Outex_TC12_001, CUReT and UIUC, and the results are shown in Table 2. It can be seen from the data in the table that m p The information entropy is about s p That is, the amount of information contained in the amplitude exceeds that of the sign. Therefore, the rational use of the amplitude can obtain more image information, which is beneficial to the texture classification results.
[0061] Table 2
[0062]
[0063] Step 3: Execute the local binary pattern algorithm based on amplitude sorting on the image. For the specific process of the algorithm processing, see steps 3a to 3d.
[0064] Step 3a, selecting a local area of a fixed scale in the texture image currently being read.
[0065] Step 3b, use bilinear interpolation to calculate the grayscale values of 8 neighborhood pixels on the circularly symmetric neighborhood set in the local area. The coordinates of the neighborhood pixels are expressed as (x+Rcos(2πp / P), y+Rsin(2πp / P)), where (x, y) are the center pixel coordinates, R is the scale, p is the index value of the current neighborhood pixel, and P is the number of neighborhood pixels.
[0066] Using Q 11 (x1,y1),Q 12 (x1,y2),Q 21 (x2,y1) and Q 22 The four gray values (x2, y2) are G(Q 11 ), G(Q 12 ), G(Q 21 ) and G(Q 22 ) are interpolated in the x-axis direction to obtain the grayscale values G(R1) and G(R2) of R1(x,y1) and R2(x,y2), and then these two grayscale values are interpolated in the y-axis to obtain the grayscale value G(P) of the desired point P(x,y). The calculation formula of the bilinear interpolation method is:
[0067]
[0068]
[0069]
[0070] Step 3c: For the central pixel of each local area, the grayscale value of the central pixel is subtracted from the grayscale value of the neighboring pixel to obtain the symbol information and amplitude information of the central pixel. The symbols are then sorted according to the amplitude. The larger the amplitude, the higher the weight is assigned to the corresponding symbol. The obtained binary pattern is the local binary pattern of the central pixel sorted based on the amplitude. For the specific implementation method, see formulas (8) and (9):
[0071]
[0072]
[0073] Wherein formula (8) is the local binary pattern of the central pixel based on the amplitude sorting, formula (9) is the function used in formula (8) to obtain the weight corresponding to each symbol, and the definition of function s(x) in formulas (8) and (9) can refer to formula (3). p and g i Both refer to the grayscale value of the neighborhood pixels, g c Refers to the grayscale value of the central pixel, and P is the number of neighborhood pixels.
[0074] Since the algorithm assigns weights based on amplitude, when the amplitudes are the same, if there is no further processing and the symbols corresponding to the amplitudes are different, the algorithm will randomly assign weights to the two symbols, which will undoubtedly bring a certain degree of randomness and uncertainty to the local binary pattern based on amplitude sorting, and thus have a certain negative impact on the classification results. Therefore, for the local binary pattern based on amplitude sorting, a specific method for dealing with this situation is given. The flowchart for obtaining weights is shown in the figure. Figure 2 As shown, it can be seen that when the amplitude is the same, no processing is performed and the next comparison cycle is directly entered, so that the weights assigned to the symbols corresponding to the same amplitude can also remain consistent, which can effectively avoid the occurrence of the above situation and further improve the effectiveness of the algorithm.
[0075] A specific example of a local binary pattern based on amplitude sorting is Figure 3 As shown in the figure, the central pixel grayscale value of 100 is used as the threshold, a local neighborhood of 3×3 is selected, and after the steps of extracting information and obtaining weights, the local binary pattern finally obtained is 01001101.
[0076] Step 3d, perform the same operation on each central pixel of the image to obtain an image processed by the local binary pattern based on amplitude sorting. Use a black and white chessboard diagram to further illustrate the problem of sorting using fixed weights, such as Figure 4As shown. It can be seen that in the image processed by the original LBP algorithm, the edge texture around the chessboard is obviously different, while the MRLBP algorithm can obtain consistent edge texture, that is, using fixed weights for sorting will enhance the texture in some directions and weaken the texture in other directions, thereby destroying the texture information of the image, which is not conducive to subsequent feature extraction. Using amplitude for sorting can solve both the problem of using fixed weights and the problem of insufficient use of amplitude information.
[0077] Step 4: pre-train the image data after the amplitude-sorted local binary pattern processing to effectively reduce the feature dimension. The specific process of pre-training is shown in steps 4a to 4d.
[0078] Step 4a, count the number of times all patterns appear in all images in the training set, that is, obtain the histogram of 256 patterns in the entire training set, and count the total number of patterns, so as to obtain the probability of each pattern appearing in the entire training set.
[0079] Step 4b, accumulate the probabilities in order from high to low, and retain the patterns with probabilities and sums within K.
[0080] Step 4c, for the modes whose probabilities are outside K and not retained, merge them into the mode closest to them, where the closest means the difference between the two modes is the smallest. The specific method is to traverse each mode and determine whether the probability corresponding to the current mode is within K. If it is within K, no processing is done and the traversal continues; if it is not within K, calculate the mode closest to the current mode with a probability within K, and merge the current mode into this mode.
[0081] In step 4d, different K values have different dimensionality reduction effects, as shown in Table 3. For different data sets, the optimal K value is also different. After testing, the optimal K value on Outex_TC10, Outex_TC12 and KTHTIPS2b is 0.8, and the optimal K value on CUReT, UMD and UIUC is 0.9. The image processed by the local binary pattern based on amplitude sorting is processed according to the pre-training rules of the texture data set to which it belongs, which can reduce the feature dimension to a certain extent and improve the accuracy of texture classification.
[0082] Table 3
[0083] K Value 1.0 0.9 0.8 0.7 0.6 0.5 Feature Dimension 256 151 109 81 61 43
[0084] Step 5, using multi-scale histograms, combining histograms of different scales in a connected manner, that is, connecting histograms of different scales in series, while ensuring low dimensionality, using more image information. The combination of histograms of different scales includes three cases: combining one-scale histogram, two-scale histogram and three-scale histogram; combining one-scale histogram, two-scale histogram, three-scale histogram and four-scale histogram; combining one-scale histogram, two-scale histogram, three-scale histogram, four-scale histogram and five-scale histogram.
[0085] After testing on three texture data sets, Outex_TC10, Outex_TC12_000 and Outex_TC12_001, the best scale selected for this implementation is a combination of one-scale histogram, two-scale histogram, three-scale histogram and four-scale histogram.
[0086] Step 6: Normalize the obtained histogram so that different features are mapped to the same interval. The method used is the Z-score normalization method, and the normalization formula is:
[0087]
[0088]
[0089]
[0090] Where x i refers to the i-th sample feature data, μ refers to the mean of all n sample feature data, δ is the standard deviation of all n sample feature data, and z is x i The normalized value.
[0091] Step 7: Use the SVM classifier to classify the texture image and obtain the classification results of each texture data set. The specific steps are: use the training sample to train the classifier to obtain the classification model; execute the classification model on the test sample to obtain the prediction result. After testing, the best parameters for this algorithm are: the kernel function selects the radial basis function RBF, the penalty factor C is 10.0, and the coefficient Gamma is 0.01.
[0092] The implementation results of the present invention are shown in Table 4 and Table 5:
[0093] Table 4 is a performance comparison of MRLBP and eight representative LBP improved algorithms, including the classification accuracy and average value on six texture datasets such as Outex_TC10, where the result of Outex_TC12 refers to the average of the results of Outex_TC12_000 and Outex_TC12_001. Table 5 is a comparison of the computational complexity between algorithms, measured by feature extraction time.
[0094] Table 4
[0095]
[0096]
[0097] As can be seen from Table 4, compared with LBP ri and LBP riu2 Compared with some of the best improved algorithms in recent years, MRLBP achieved similar texture classification accuracy on these six datasets, with an average accuracy only about 2% lower than the best result, achieving good results. Among them, due to too many changes in illumination, angle and scale between the training set and the test set of the KTHTIPS2b dataset, the classification accuracy of each algorithm on this dataset was not high, and the effect of MRLBP was second only to MRELBP and higher than the other seven algorithms.
[0098] Table 5
[0099] Operator <![CDATA[LBP ri ]]> <![CDATA[LBP riu2 ]]> <![CDATA[LTP riu2 ]]> CLBP disCLBP CINIRD MRELBP MRLBP Feature extraction time t 0.98t 2.66t 1.47t 6.72t 1.31t 19.73t 1.32t
[0100] It can be seen from Table 5 that LBP ri The computational complexity of MRLBP is used as a benchmark to calculate and compare the feature extraction time under a single scale. The computational complexity of MRLBP is 1.32t, which is only slightly higher than that of CINIRD algorithm, but lower than that of several other improved algorithms and higher than that of LTP. riu2 , the computational complexity of disCLBP and MRELBP algorithms is several times lower.
[0101] The implementation results of the present invention show that the texture classification method based on MRLBP has achieved good implementation results and maintained the advantage of low computational complexity of the original local binary pattern. In addition, the texture classification method based on MRLBP has good rotation invariance and certain noise robustness because of its idea based on amplitude sorting.
[0102] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.
Claims
1. A texture classification method based on MRLBP, characterized in that: The following steps are involved: Step 1, read the image in the texture data set, if the image is not a grayscale image, convert it into a grayscale image first; Step 2, processing the image by using the amplitude-sorted local binary pattern algorithm, and obtaining an image processed by the amplitude-sorted local binary pattern (MRLBP) algorithm by using the sign information and the amplitude information; The implementation of the amplitude sorting based local binary pattern algorithm described in step 2 comprises the following steps: Step 21, selecting a local area of a suitable scale in the texture image currently read; Step 22, using bilinear interpolation method, calculate the grayscale values of P neighborhood pixels of the circularly symmetric neighborhood set in the local area, where the coordinates of the neighborhood pixels are expressed as (x+Rcos(2πp / P), y+Rsin(2πp / P)), (x, y) are the coordinates of the center pixel, R is the scale, p is the index value of the current neighborhood pixel, and P is the number of neighborhood pixels; Step 23, for the central pixel of each local area, the gray value of the central pixel is subtracted from the gray value of the neighboring pixel to obtain the symbol information and amplitude information of the central pixel; then the symbols are sorted according to the amplitude, that is, different weights are assigned. The larger the amplitude, the higher the weight assigned to the corresponding symbol. The obtained binary pattern is the local binary pattern of the central pixel sorted based on the amplitude. For the specific implementation method, see formulas (8) and (9): in, Formula (8) is the local binary pattern of the central pixel based on the amplitude sorting, and Formula (9) is the function used to obtain the weight corresponding to each symbol in Formula (8), where g p and g i Both refer to the grayscale value of the neighborhood pixels, g c Refers to the gray value of the center pixel; Step 24, performing the same operation on each central pixel of the image to obtain an image processed by the local binary pattern based on amplitude sorting; Step 3, extracting features from the processed image, first pre-training the image data after the amplitude-sorted local binary pattern processing to effectively reduce the feature dimension; Step 4: Use multi-scale histograms to combine histograms of different scales in a connected manner to obtain higher classification accuracy; Step 5: finally, the obtained histogram is normalized to remove the dimension of the feature, so that different features are mapped to the same interval and the convergence speed of the classifier is improved; Step 6: Use SVM classifier to classify the texture image.
2. A texture classification method based on MRLBP according to claim 1, characterized in that: The calculation formula of the bilinear interpolation method in step 22 is: That is, using Q 11 (x1,y1),Q 12 (x1,y2),Q 21 (x2,y1) and Q 22 The four gray values (x2, y2) are G(Q 11 ), G(Q 12 ), G(Q 21 ) and G(Q 22 ) are interpolated in the x-axis direction to obtain the grayscale values G(R1) and G(R2) of R1(x,y1) and R2(x,y2), and then these two grayscale values are interpolated in the y-axis to obtain the grayscale value G(P) of the required point P(x,y).
3. A texture classification method based on MRLBP according to claim 1, characterized in that: The pre-training described in step 3 includes the following steps: Step 31, counting the probability of each pattern appearing in the entire texture data set, that is, obtaining a histogram of 256 patterns on the entire training set, and counting the total number of patterns, thereby obtaining the probability of each pattern appearing in the entire training set; Step 32, add the probabilities from high to low, and keep the patterns whose probabilities and sums are within K; Step 33, for the mode whose probability is not within K, merge it into the mode that is closest to it; the closest means that the difference between the two modes is the smallest. The specific method is to traverse each mode and determine whether the probability corresponding to the current mode is within K. If it is within K, no processing is performed and the traversal continues; if it is not within K, calculate the mode that is closest to the current mode with a probability within K, and merge the current mode into this mode; Step 34, processing the image after the amplitude-sorted local binary pattern processing according to the pre-training rules of the corresponding texture data set can reduce the feature dimension to a certain extent.
4. The texture classification method based on MRLBP according to claim 1, characterized in that: The combination of histograms of different scales includes the following three cases: Combining one-scale histogram, two-scale histogram and three-scale histogram; Combination of one-scale histogram, two-scale histogram, three-scale histogram and four-scale histogram; The one-scale histogram, the two-scale histogram, the three-scale histogram, the four-scale histogram and the five-scale histogram are combined.
5. The texture classification method based on MRLBP according to claim 1, characterized in that: The P value of the P neighborhood pixels in step 22 is 8, and the angles of these 8 neighborhood pixels on the circularly symmetric neighborhood set are 0°, 45°, 90°, 135°, 180°, 225°, 270° and 315° respectively.
6. The texture classification method based on MRLBP according to claim 1, characterized in that: When weights are assigned to symbols according to the amplitude in step 23, if the amplitudes are the same, the symbols with the same amplitude are assigned the same weights.
7. The texture classification method based on MRLBP according to claim 2, characterized in that: The probability and K include five values: 0.5, 0.6, 0.7, 0.8, 0.9 and 1.
0.
8. The texture classification method based on MRLBP according to claim 1, characterized in that: In step 5, the Z-score normalization method is used to normalize the histogram.
9. The texture classification method based on MRLBP according to claim 1, characterized in that: In step 6, the SVM classifier kernel function selects the radial basis function RBF, the penalty factor C is 10.0, and the coefficient Gamma is 0.01.
