CN107239783B - Coal rock identification method based on extended local binary pattern and regression analysis - Google Patents

Abstract

The invention discloses a coal rock identification method based on an extended local binary pattern and regression analysis. In a sample training stage, obtaining a plurality of coal sample sub-images and rock sample sub-images through a series of processing such as camera shooting, sub-image interception, graying and the like, extracting expansion local binary pattern feature vectors of each sub-image, carrying out regression analysis on the feature vectors and class labels thereof to obtain a feature optimal combination, and obtaining a training sample optimal feature vector based on the feature optimal combination; in the coal rock identification stage, an unknown class sample sub-image is obtained through a series of processing, an extended local binary pattern feature vector of the unknown class sample sub-image is extracted, and an unknown class sample preferred feature vector is obtained based on a feature preferred combination in the sample training stage. And judging the coal rock category to which the unknown class sample belongs by comparing the preferred characteristic vectors between the unknown class sample and the training sample. The method has obvious advantages in the aspects of flexibility, reliability, instantaneity and the like.

Description

Coal rock identification method based on extended local binary pattern and regression analysis

Technical Field

The invention relates to a coal rock identification method based on an extended local binary pattern and regression analysis, and belongs to the technical field of image identification.

Background

Coal and rock identification means that coal and rock are automatically distinguished by various technical means. In the process of coal resource mining and transportation, a plurality of production links need to distinguish coal and rock, such as height adjustment of a roller of a coal mining machine, control of a fully mechanized top coal caving process, selection of gangue from raw coal of a coal preparation plant and the like. Since the 50 s of the 20 th century, a series of researches on coal rock identification methods in major coal-producing countries in the world such as south africa, australia, germany, the united states and china are carried out, and some representative research results such as a natural gamma ray detection method, a radar detection method, an infrared detection method, an active power detection method, a vibration signal detection method, a sound signal detection method and the like are generated successively. However, these methods all have the following common problems: (1) various sensors need to be installed and deployed on the existing equipment, and related devices are complex in structure and high in manufacturing cost; (2) mechanical equipment such as a coal mining machine, a heading machine and the like is complex in stress, severe in vibration and severe in abrasion in the coal production process, a sensor is relatively difficult to deploy, an electronic circuit of the sensor is easy to damage, and the reliability of the device is poor; (3) for different types of mechanical carrier equipment, great differences exist between the type selection of the sensor and the selection of the installation position, and personalized customization is needed, so that the universality is poor.

Through observation of blocky coal and rock samples, the coal and the rock have great differences in color, luster, texture and the like. When the coal and the rock are imaged by the existing digital camera, the visual information of the coal and the rock is necessarily hidden in the acquired digital image, so that the coal and the rock are distinguished by mining the visual information in the digital image of the coal and the rock. The existing coal and rock identification method based on image processing has a larger promotion space in the aspects of robustness, identification rate and the like.

Disclosure of Invention

In order to overcome the defects of the existing coal rock identification method, the invention provides the coal rock identification method based on the extended local binary pattern and the regression analysis, and the method has the advantages of strong real-time performance, high identification rate, good robustness and the like, and is beneficial to improving the production efficiency and the safety degree of the modern coal mine.

The coal rock identification method is realized by adopting the following technical scheme, comprises a sample training stage and a coal rock identification stage, and specifically comprises the following steps:

RS1, in a sample training stage, acquiring m coal sample images and m rock sample images, intercepting subgraphs without non-coal rock background and carrying out gray processing on the subgraphs, and respectively recording the processed coal sample subgraphs and rock sample subgraphs as c₁,c₂,…,c_mAnd s₁,s₂,…,s_m；

Rs2. set sampling radius r 2, radial spacing 2, sampling neighborhood number p 8, and sliding window size ω for median filtering_rCount c for 3 × 3₁,c₂,…,c_mAnd s₁,s₂,…,s_mWith median filtering and with rotationRobust extended local binary pattern 3-dimensional joint histogram with invariant characteristics and uniform characteristics and formed by combining 3 descriptors, namely descriptor based on central point intensity, descriptor based on neighborhood point intensity and descriptor based on radial difference

And

RS3. respectively handle

And

transformed into a 1-dimensional histogram, and then normalized to obtain a feature row vector α₁,α₂,…,α_m∈R^1×200And β₁,β₂,…,β_m∈R^1×200；

Rs4. set sampling radius r 4, radial spacing 2, sampling neighborhood number p 8, and sliding window size ω for median filtering_rCount c for 5 × 5₁,c₂,…,c_mAnd s₁,s₂,…,s_mThe robust extended local binary pattern 3-dimensional joint histogram which adopts median filtering and has rotation invariant characteristic and uniform characteristic and is formed by combining 3 descriptors of descriptor based on central point intensity, descriptor based on neighborhood point intensity and descriptor based on radial difference

And

RS5. respectively

And

transformed into a 1-dimensional histogram, and then normalized to obtain a feature row vector η₁,η₂,…,η_m∈R^1×200And mu₁,μ₂,…,μ_m∈R^1×200；

Rs6. set sampling radius r 6, radial spacing 2, sampling neighborhood number p 8, and sliding window size ω for median filtering_rC is counted as 7 × 7₁,c₂,…,c_mAnd s₁,s₂,…,s_mThe robust extended local binary pattern 3-dimensional joint histogram which adopts median filtering and has rotation invariant characteristic and uniform characteristic and is formed by combining 3 descriptors of descriptor based on central point intensity, descriptor based on neighborhood point intensity and descriptor based on radial difference

And

RS7. respectively handle

And

transforming into 1-dimensional histogram, and normalizing them to obtain characteristic row vector theta₁,θ₂,…,θ_m∈R^1×200And

rs8. set sampling radius r 8, radial spacing 2, sampling neighborhood number p 8, and sliding window size ω for median filtering_rCount c for 9 × 9₁,c₂,…,c_mAnd s₁,s₂,…,s_mThe robust extended local binary pattern 3-dimensional joint histogram which adopts median filtering and has rotation invariant characteristic and uniform characteristic and is formed by combining 3 descriptors of descriptor based on central point intensity, descriptor based on neighborhood point intensity and descriptor based on radial difference

And

RS9. respectively

And

transformed into a 1-dimensional histogram, and then normalized to obtain a feature row vector k₁,κ₂,…,κ_m∈R^1×200And upsilon₁,υ₂,…,υ_m∈R^1×200；

RS10. construction of c separately₁,c₂,…,c_mAnd s₁,s₂,…,s_mOriginal feature column vector x₁＝[α₁,η₁,θ₁,κ₁]^T,x₂＝[α₂,η₂,θ₂,κ₂]^T,…,x_m＝[α_m,η_m,θ_m,κ_m]^T∈R^800×1And

，…,

wherein T is a transposition operation;

RS11. constructing an original characteristic matrix X ═ X of the coal petrography training sample₁,x₂,…,x_m,x_m+1,x_m+2,…,x_2m]^T∈R^{2m ×800}And its category label matrix Y ∈ R^2m×2Wherein T is a transpose operation, the first m rows and 1 st columns of data padding of Y is 1, the first m rows and 2 nd columns of data padding of Y is 0, the last m rows and 1 st columns of data padding of Y is 0, and the last m rows and 2 nd columns of data padding of Y is 1;

RS12, setting a regularization parameter lambda, iteration times K and an optimal feature number d, and performing regression analysis on X and Y to obtain d column marks j about X and beneficial to distinguishing coal rocks₁,j₂,…,j_dWherein j is not less than 1₁,j₂,…,j_d≤800；

RS13. sequentially extracting j' th of X₁,j₂,…,j_dArranging the columns, and then forming a final characteristic matrix of the coal rock training sample according to the arrangement of the columns

Wherein X (col. j)₁),X(col·j₂),…,X(col·j_d) J' th of X respectively₁,j₂,…,j_dColumns;

in the coal rock identification stage, acquiring an unknown class sample image, intercepting a sub-image without a non-coal rock background and carrying out gray processing on the sub-image, wherein the processed unknown class sub-image is marked as q;

RS15, setting the sampling radius r to be 2, the radial interval to be 2, the sampling neighborhood number p to be 8 and filtering the medianSliding window size omega of wave_r3 × 3, the robust extended local binary pattern 3-dimensional joint histogram H with the combination of 3 descriptors, namely a descriptor based on central point intensity, a descriptor based on neighborhood point intensity and a descriptor based on radial difference, which adopts median filtering and has rotation invariant characteristic and uniform characteristic_q；

RS16. A. H_qTransforming into 1-dimensional histogram, and normalizing to obtain characteristic row vector_q∈R^{1 ×200}；

Rs17. set sampling radius r 4, radial spacing 2, sampling neighborhood number p 8, and sliding window size ω for median filtering_r5 × 5, the robust extended local binary pattern 3-dimensional joint histogram Z with the rotation invariant characteristic and the uniform characteristic and adopting the median filtering is combined by 3 descriptors of a descriptor based on the central point intensity, a descriptor based on the neighborhood point intensity and a descriptor based on the radial difference_q；

RS18. handle Z_qThe shape is changed into a 1-dimensional histogram, and then the histogram is normalized to obtain a characteristic row vector rho_q∈R^{1 ×200}；

Rs19. set sampling radius r 6, radial spacing 2, sampling neighborhood number p 8, and sliding window size ω for median filtering_r7 × 7, the robust extended local binary pattern 3-dimensional joint histogram F with the combination of 3 descriptors, namely a descriptor based on central point intensity, a descriptor based on neighborhood point intensity and a descriptor based on radial difference, which adopts median filtering and has rotation invariant characteristic and uniform characteristic, of statistics q_q；

RS20. A. F_qThe shape is changed into a 1-dimensional histogram, and then the histogram is normalized to obtain a characteristic row vector sigma_q∈R^{1 ×200}；

Rs21. set sampling radius r 8, radial spacing 2, sampling neighborhood number p 8, and sliding window size ω for median filtering_r9 × 9, statistical q center-based with median filtering and with rotation invariant and uniform propertiesPoint intensity descriptor, neighborhood point intensity-based descriptor and radial difference-based descriptor, namely 3 descriptors are combined to form robust extended local binary pattern 3-dimensional joint histogram T_q；

RS22. separately treating T_qTransformed into a 1-dimensional histogram, which is then normalized to obtain a feature row vector ξ_q∈R^1×200；

RS23. construction of the raw feature row vector x of q_q＝[_q,ρ_q,σ_q,ξ_q]∈R^1×800；

RS24. successively extracting x_qJ (d) of₁,j₂,…,j_dElements, then arranging the final feature row vector constituting q in order

Wherein x_q(j₁),x_q(j₂),…,x_q(j_d) Are respectively x_qJ (d) of₁,j₂,…,j_dAn element;

RS25. by calculating

Determining the coal rock class of q, if satisfied

Then judging q as coal; otherwise, q is determined to be rock, where i is

I is 1,2, …,2m, j is

Column symbol of

The element numbers of (1), (2), (…), (d),

of the elements located in row i and column j,

is composed of

The jth element of (1).

The regression analysis described in step RS12 includes the following steps:

RS1201. construct symbol matrix U ∈ R^2m×2Then initializing m elements positioned in the first m rows and the 1 st column in the U by using 1, initializing m elements positioned in the first m rows and the 2 nd column in the U by using-1, initializing m elements positioned in the second m rows and the 1 st column in the U by using-1, and initializing m elements positioned in the second m rows and the 2 nd column in the U by using 1;

RS1202. construct data matrix X⁰∈R^2m×800Then by

Column by column initialization X⁰Wherein τ is X⁰And X, τ ═ 1,2, …,800,

is X⁰Column τ of, x_τColumn τ of X, mean (X)_τ) Is x_τMean of all elements in;

RS1203. respectively constructing weight matrix W_old,W_new∈R^800×2And an offset vector b_old,b_new∈R^2×1；

RS1204, first pass W_old＝(X⁰)^T{[X⁰(X⁰)^T+λI_2m]^-1Y initialization W_oldThen through W_new＝W_oldInitializing W_newWhere T and-1 are the transposition and inversion operations, respectively, I_2mIs an identity matrix of 2m order;

RS1205. constructing a data matrix L_b∈R^2×2mThen by

Initialization L_b；

RS1206. first pass

Initialization b_oldThen through b_new＝b_oldInitialization b_newWherein L is_b(col. j) is L_bJ is L_bThe list of (1), (2), (…), (2 m);

RS1207, respectively constructing data matrixes A, B, G ∈ R^2m×2Then initializing them to a zero matrix;

RS1208. constructing data matrix X¹∈R^2m×801Then initialize X with X¹First 800 columns of (A), X¹All elements of column 801 are initialized to 1000, i.e., X¹＝[X,1000e_2m]∈R^2m×801Wherein e is_2m∈R^2m×1A column vector with all elements 1;

RS1209. defining iteration number k₁And is initialized to 0;

RS1210. by

Update A, where T is a transpose operation, e_2m∈R^2m×1A column vector with all elements 1;

rs1211, first updating B by B ═ U ═ a, and then replacing the negative element in B with 0, where ≥ is a Hadamard product operation of the matrix;

rs1212. update G by G ═ B) + Y, wherein ═ Hadamard product operation of the matrix;

RS1213. describe mathematics as

The optimization problem (W) is recorded as KeyProblem, and the solution W of the KeyProblem is obtained₀∈R^801×2Wherein | · | purple light_2,1Is a matrix of_2,1Norm in matrix Ψ ∈ R^u0×v0For the purpose of example only,

i0 is a row indicator of Ψ, i0 ═ 1,2, …, u0, j0 is a column indicator of Ψ, j0 ═ 1,2, …, v0, Ψ (i0, j0) is an element in Ψ located in the j0 th column of the i0 th row, and u0 and v0 are the row number and column number of Ψ, respectively;

RS1214. with W₀Update W on the first 800 rows_new；

RS1215. through b_new＝1000×[W₀(row·801)]^TUpdate b_newWhere T is a transposition operation, W₀(row 801) is W₀Line 801;

RS1216. iteration number k₁Self-increment by 1;

RS1217. if satisfied at the same time

And k₁<K, then use W_newUpdate W of the value of_oldBy b_newUpdate of value b_oldThen, step RS 1210-RS 1217 is performed, where | · | | | ceiling_FAnd | · | non-conducting phosphor₂The Frobenius norm of the matrix and the 2-norm of the vector are respectively; otherwise, executing step RS 1218;

RS1218. constructing a weight square matrix W₂∈R^800×2Then through W₂＝W_new⊙W_newInitializing W₂Wherein ⊙ is a Hadamard product operation of the matrix;

RS1219. construction of weight vector

Then pass through

Updating

Wherein W₂(col.1) and W₂(col. 2) are each W₂Column 1 and column 2;

RS1220. finish regression analysis, return

The sequence number j of the first d elements in the sequence from big to small₁,j₂,…,j_dAs d columns for X.

The solution of keypromlem in step RS1213 includes the following steps:

RS121301. construction of data matrix X₊∈R^2m×(801+2m)Then with X¹Initialization of X₊First 801 columns of (1), with 2m order identity matrix I_2mInitialization of X₊The last 2m column of (i.e. X)₊＝[X¹,I_2m]∈R^2m×(801+2m)In which I_2mIs an identity matrix of 2m order;

RS121302. construction of diagonal element vector φ ∈ R^(801+2m)×1Then all elements of phi are initialized to 1;

RS121303. defining iteration number k₂And initialized to 0, define and variable Sum and initialized to-1000.000;

RS121304. construction of diagonal matrix Λ∈ R^{(801+2m)×(801+2m)}；

Rs121305. via Λ_jj＝φ_jJ is 1,2, …, (801+2m) sequentially updates the main diagonal element of Λ, where j is the element number of phi and the main diagonal element number of Λ, phi_jIs the jth element of phi, Λ_jjThe jth main diagonal element of Λ, i.e., Λ_jjΛ for the element located in row jth column jth;

RS121306. separately construct the data matrix W₀∈R^801×2And W₃,W₄∈R^(801+2m)×2Then initializing them to a zero matrix;

RS121307. by

Updating W₃Wherein T and-1 are transposition operation and inversion operation respectively;

RS121308. with W₄＝W₃⊙W₃Update W of the calculation result of (2)₄Wherein ⊙ is a Hadamard product operation of the matrix;

rs121309. first pass through ═ W₄(col·1)+W₄(col.2) updating phi, if zero element exists in phi, replacing zero element with 0.000001, wherein W₄(col.1) and W₄(col. 2) are each W₄Column 1 and column 2;

rs121310. iteration number k₂Self-increment by 1;

RS121311. if satisfied at the same time

And k₂<K, then pass

Update Sum and then perform steps RS 121305-RS 121311, where j is the element number of φ, φ_jThe jth element of phi; otherwise, go to step RS 121312;

RS121312. with W₃The first 801 rows of update W₀；

RS121313. complete the solution of KeyProblem, return the solution W of KeyProblem₀。

Drawings

FIG. 1 is a basic flow diagram of a coal-rock identification method based on extended local binary patterns and regression analysis;

FIG. 2 is a basic flow diagram of the regression analysis of the present invention;

FIG. 3 is a solution W for KeyProblem according to the present invention₀A basic flow diagram of (1);

Detailed Description

On the basis of carrying out experimental analysis on images of main coal types and rock types in Henan, Shanxi, Shaanxi and other places of China, the invention provides a coal and rock identification method based on an extended local binary pattern and regression analysis.

The invention is described in further detail below with reference to the figures and the detailed description.

Referring to fig. 1, the coal rock identification method based on the extended local binary pattern and the regression analysis specifically includes the following steps:

SS1, in a sample training stage, acquiring m coal sample images and m rock sample images, intercepting subgraphs not containing non-coal rock backgrounds and carrying out gray processing on the subgraphs, and respectively recording the processed coal sample subgraphs and rock sample subgraphs as c₁,c₂,…,c_mAnd s₁,s₂,…,s_m；

Ss2. set sampling radius r 2, radial spacing 2, sampling neighborhood number p 8, and sliding window size ω for median filtering_rCount c for 3 × 3₁,c₂,…,c_mAnd s₁,s₂,…,s_mThe robust extended local binary pattern 3-dimensional joint histogram which adopts median filtering and has rotation invariant characteristic and uniform characteristic and is formed by combining 3 descriptors of descriptor based on central point intensity, descriptor based on neighborhood point intensity and descriptor based on radial difference

And

SS3. respectively handle

And

transformed into a 1-dimensional histogram, and then normalized to obtain a feature row vector α₁,α₂,…,α_m∈R^1×200And β₁,β₂,…,β_m∈R^1×200；

SS4. set sampling radius r 4, radial spacing 2, sampling neighborhood number p 8, and sliding window size ω for median filtering_rCount c for 5 × 5₁,c₂,…,c_mAnd s₁,s₂,…,s_mThe robust extended local binary pattern 3-dimensional joint histogram which adopts median filtering and has rotation invariant characteristic and uniform characteristic and is formed by combining 3 descriptors of descriptor based on central point intensity, descriptor based on neighborhood point intensity and descriptor based on radial difference

And

SS5. respectively handle

And

transformed into a 1-dimensional histogram, and then normalized to obtain a feature row vector η₁,η₂,…,η_m∈R^1×200And mu₁,μ₂,…,μ_m∈R^1×200；

SS6. set sampling radius r 6, radial spacing 2, sampling neighborhood number p 8, and sliding window size ω for median filtering_rC is counted as 7 × 7₁,c₂,…,c_mAnd s₁,s₂,…,s_mThe robust extended local binary pattern 3-dimensional joint histogram which adopts median filtering and has rotation invariant characteristic and uniform characteristic and is formed by combining 3 descriptors of descriptor based on central point intensity, descriptor based on neighborhood point intensity and descriptor based on radial difference

And

SS7. respectively handle

And

transforming into 1-dimensional histogram, and normalizing them to obtain characteristic row vector theta₁,θ₂,…,θ_m∈R^1×200And

ss8. set sampling radius r 8, radial spacing 2, sampling neighborhood number p 8, and sliding window size ω for median filtering_rCount c for 9 × 9₁,c₂,…,c_mAnd s₁,s₂,…,s_mThe robust extended local binary pattern 3-dimensional joint histogram which adopts median filtering and has rotation invariant characteristic and uniform characteristic and is formed by combining 3 descriptors of descriptor based on central point intensity, descriptor based on neighborhood point intensity and descriptor based on radial difference

And

SS9. respectively handle

And

transformed into a 1-dimensional histogram, and then normalized to obtain a feature row vector k₁,κ₂,…,κ_m∈R^1×200And upsilon₁,υ₂,…,υ_m∈R^1×200；

SS10. construction of c separately₁,c₂,…,c_mAnd s₁,s₂,…,s_mOriginal feature column vector x₁＝[α₁,η₁,θ₁,κ₁]^T,x₂＝[α₂,η₂,θ₂,κ₂]^T,…,x_m＝[α_m,η_m,θ_m,κ_m]^T∈R^800×1And

，…,

wherein T is a transposition operation;

SS11. constructing an original characteristic matrix X ═ X of the coal rock training sample₁,x₂,…,x_m,x_m+1,x_m+2,…,x_2m]^T∈R^{2m ×800}And its category label matrix Y ∈ R^2m×2Wherein T is a transpose operation, the first m rows and 1 st columns of data padding of Y is 1, the first m rows and 2 nd columns of data padding of Y is 0, the last m rows and 1 st columns of data padding of Y is 0, and the last m rows and 2 nd columns of data padding of Y is 1;

SS12, setting a regularization parameter lambda, iteration times K and an optimal feature number d, and performing regression analysis on X and Y to obtain d column marks j about X and beneficial to distinguishing coal rocks₁,j₂,…,j_dWherein j is not less than 1₁,j₂,…,j_d≤800；

SS13. extract j of X in sequence₁,j₂,…,j_dArranging the columns, and then forming a final characteristic matrix of the coal rock training sample according to the arrangement of the columns

X(col·j₂),…,X(col·j_d)]∈R^2m×dWherein X (col. j)₁),X(col·j₂),…,X(col·j_d) J' th of X respectively₁,j₂,…,j_dColumns;

SS14, in the coal rock identification stage, collecting an unknown class sample image, intercepting a sub-image without a non-coal rock background and carrying out gray processing on the sub-image, wherein the processed unknown class sub-image is marked as q;

ss15. set sampling radius r 2, radial spacing 2, sampling neighborhood number p 8, and sliding window size ω for median filtering_r3 × 3, the robust extended local binary pattern 3-dimensional joint histogram H with the combination of 3 descriptors, namely a descriptor based on central point intensity, a descriptor based on neighborhood point intensity and a descriptor based on radial difference, which adopts median filtering and has rotation invariant characteristic and uniform characteristic_q；

SS16. treating H_qTransforming into 1-dimensional histogram, and normalizing to obtain characteristic row vector_q∈R^{1 ×200}；

Ss17. set the sampling radius r 4, the radial spacing 2, the number of sampling neighborhoods p 8, and the sliding window size ω for median filtering_r5 × 5, the robust extended local binary pattern 3-dimensional joint histogram Z with the rotation invariant characteristic and the uniform characteristic and adopting the median filtering is combined by 3 descriptors of a descriptor based on the central point intensity, a descriptor based on the neighborhood point intensity and a descriptor based on the radial difference_q；

SS18. handle Z_qThe shape is changed into a 1-dimensional histogram, and then the histogram is normalized to obtain a characteristic row vector rho_q∈R^{1 ×200}；

Ss19. set sampling radius r 6, radial spacing 2, sampling neighborhood number p 8, and sliding window size ω for median filtering_r7 × 7, statistics q using median filtering and having rotation invariant and uniform properties are based on central point intensity descriptors, neighborhood point intensity descriptors and radial differenceRobust extended local binary pattern 3-dimensional joint histogram F formed by combining 3 descriptors of sub-descriptors_q；

SS20. A. B_qThe shape is changed into a 1-dimensional histogram, and then the histogram is normalized to obtain a characteristic row vector sigma_q∈R^{1 ×200}；

Ss21. set sampling radius r 8, radial spacing 2, sampling neighborhood number p 8, and sliding window size ω for median filtering_r9 × 9, the robust extended local binary pattern 3-dimensional joint histogram T with rotation invariant and uniform characteristics and adopting median filtering and with the rotation invariant characteristic and the uniform characteristic is formed by combining 3 descriptors, namely a descriptor based on central point intensity, a descriptor based on neighborhood point intensity and a descriptor based on radial difference_q；

SS22. separately treating T_qTransformed into a 1-dimensional histogram, which is then normalized to obtain a feature row vector ξ_q∈R^1×200；

SS23. construct the raw feature row vector x for q_q＝[_q,ρ_q,σ_q,ξ_q]∈R^1×800；

SS24. successively extract x_qJ (d) of₁,j₂,…,j_dElements, then arranging the final feature row vector constituting q in order

Wherein x_q(j₁),x_q(j₂),…,x_q(j_d) Are respectively x_qJ (d) of₁,j₂,…,j_dAn element;

SS25. by calculation

Determining the coal rock class of q, if satisfied

Then judging q as coal; otherwise, q is determined to be rock, where i is

I is 1,2, …,2m, j is

Column symbol of

The element numbers of (1), (2), (…), (d),

is composed of

Of the elements located in row i and column j,

is composed of

The jth element of (1).

Referring to FIG. 2, the regression analysis described in step SS12 includes the following steps:

SS1201. construct the symbol matrix U ∈ R^2m×2Then initializing m elements positioned in the first m rows and the 1 st column in the U by using 1, initializing m elements positioned in the first m rows and the 2 nd column in the U by using-1, initializing m elements positioned in the second m rows and the 1 st column in the U by using-1, and initializing m elements positioned in the second m rows and the 2 nd column in the U by using 1;

SS1202. construct data matrix X⁰∈R^2m×800Then by

Column by column initialization X⁰Wherein τ is X⁰And X, τ ═ 1,2, …,800,

is X⁰Column τ of, x_τColumn τ of X, mean (X)_τ) Is x_τMean of all elements in;

SS1203. respectively constructing weight matrices W_old,W_new∈R^800×2And an offset vector b_old,b_new∈R^2×1；

SS1204, first pass W_old＝(X⁰)^T{[X⁰(X⁰)^T+λI_2m]^-1Y initialization W_oldThen through W_new＝W_oldInitializing W_newWhere T and-1 are the transposition and inversion operations, respectively, I_2mIs an identity matrix of 2m order;

SS1205. construct the data matrix L_b∈R^2×2mThen by

Initialization L_b；

SS1206. first pass

Initialization b_oldThen through b_new＝b_oldInitialization b_newWherein L is_b(col. j) is L_bJ is L_bThe list of (1), (2), (…), (2 m);

SS1207. construction of data matrix A, B, G ∈ R separately^2m×2Then initializing them to a zero matrix;

SS1208. construct data matrix X¹∈R^2m×801Then initialize X with X¹First 800 columns of (A), X¹All elements of column 801 are initialized to 1000, i.e., X¹＝[X,1000e_2m]∈R^2m×801Wherein e is_2m∈R^2m×1A column vector with all elements 1;

SS1209. define iteration number k₁And is initialized to 0;

SS1210. by

Update A, where T is a transpose operation, e_2m∈R^2m×1A column vector with all elements 1;

SS1211 updating B by first updating B by B ═ U ≧ A, and then replacing the negative element in B by 0, wherein | _ is Hadamard product operation of the matrix;

ss1212 updating G by G ═ B) + Y, wherein ═ Hadamard product operation of the matrix;

SS1213. describe mathematics as

The optimization problem (W) is recorded as KeyProblem, and the solution W of the KeyProblem is obtained₀∈R^801×2Wherein | · | purple light_2,1Is a matrix of_2,1Norm in matrix Ψ ∈ R^u0×v0For the purpose of example only,

i0 is a row indicator of Ψ, i0 ═ 1,2, …, u0, j0 is a column indicator of Ψ, j0 ═ 1,2, …, v0, Ψ (i0, j0) is an element in Ψ located in the j0 th column of the i0 th row, and u0 and v0 are the row number and column number of Ψ, respectively;

SS1214 with W₀Update W on the first 800 rows_new；

SS1215. through b_new＝1000×[W₀(row·801)]^TUpdate b_newWhere T is a transposition operation, W₀(row 801) is W₀Line 801;

SS1216. iteration number k₁Self-increment by 1;

SS1217. if satisfied at the same time

And k₁<K, then use W_newUpdate W of the value of_oldBy b_newUpdate of value b_oldThen, step SS 1210-SS 1217 is performed, in which | · | | | ceiling_FAnd | · | non-conducting phosphor₂The Frobenius norm of the matrix and the 2-norm of the vector are respectively; otherwise, go to step SS 1218;

SS1218. construct the weight-squared matrix W₂∈R^800×2Then through W₂＝W_new⊙W_newInitializing W₂Which isThe middle ⊙ is the Hadamard product operation of the matrix;

SS1219. construction of weight vectors

Then pass through

Updating

Wherein W₂(col.1) and W₂(col. 2) are each W₂Column 1 and column 2;

SS1220. complete regression analysis, return

The sequence number j of the first d elements in the sequence from big to small₁,j₂,…,j_dAs d columns for X.

Referring to FIG. 3, the solution W of KeyProblem described in step SS1213 is obtained₀The method comprises the following specific steps:

SS121301. construction of data matrix X₊∈R^2m×(801+2m)Then with X¹Initialization of X₊First 801 columns of (1), with 2m order identity matrix I_2mInitialization of X₊The last 2m column of (i.e. X)₊＝[X¹,I_2m]∈R^2m×(801+2m)In which I_2mIs an identity matrix of 2m order;

SS121302. construction of diagonal element vector φ ∈ R^(801+2m)×1Then all elements of phi are initialized to 1;

SS121303. defining iteration number k₂And initialized to 0, define and variable Sum and initialized to-1000.000;

SS121304. construction of diagonal matrix Λ∈ R^{(801+2m)×(801+2m)}；

SS121305. by Λ_jj＝φ_jJ is 1,2, …, (801+2m) sequentially updates the main diagonal element of Λ, where j is the element number of phi and the main diagonal element number of Λ, phi_jIs the jth element of phi, Λ_jjThe jth main diagonal element of Λ, i.e., Λ_jjΛ for the element located in row jth column jth;

SS121306. separately construct the data matrix W₀∈R^801×2And W₃,W₄∈R^(801+2m)×2Then initializing them to a zero matrix;

SS121307. by

Updating W₃Wherein T and-1 are transposition operation and inversion operation respectively;

SS121308. with W₄＝W₃⊙W₃Update W of the calculation result of (2)₄Wherein ⊙ is a Hadamard product operation of the matrix;

ss121309. first pass phi ═ W₄(col·1)+W₄(col.2) updating phi, if zero element exists in phi, replacing zero element with 0.000001, wherein W₄(col.1) and W₄(col. 2) are each W₄Column 1 and column 2;

SS121310. iteration number k₂Self-increment by 1;

SS121311. if satisfied at the same time

And k₂<K, then pass

Update Sum and then perform steps SS 121305-SS 121311, where j is the element number of φ, φ_jThe jth element of phi; otherwise, go to step SS 121312;

SS121312. with W₃The first 801 rows of update W₀；

SS121313. the solution of KeyProblem is completed, and the solution W of KeyProblem is returned₀。

It should be noted that the above-mentioned embodiment examples are used to further illustrate the present invention, and the embodiment examples should not be construed as limiting the scope of the present invention.

Claims (3)

1. The coal rock identification method based on the extended local binary pattern and the regression analysis is characterized by comprising the following steps of:

QS1, in the stage of sample training, acquiring m coal sample images and m rock sample images, intercepting subgraphs without non-coal rock background and carrying out gray processing on the subgraphs, and respectively recording the processed coal sample subgraphs and rock sample subgraphs as c₁,c₂,…,c_mAnd s₁,s₂,…,s_m；

Qs2. set the sampling radius r 2, the radial spacing 2, the number of sampling neighborhoods p 8, and the sliding window size ω for median filtering_rCount c for 3 × 3₁,c₂,…,c_mAnd s₁,s₂,…,s_mThe robust extended local binary pattern 3-dimensional joint histogram which adopts median filtering and has rotation invariant characteristic and uniform characteristic and is formed by combining 3 descriptors of descriptor based on central point intensity, descriptor based on neighborhood point intensity and descriptor based on radial difference

And

QS3. respectively handle

And

transformed into a 1-dimensional histogram, and then normalized to obtain a feature row vector α₁,α₂,…,α_m∈R^1×200And β₁,β₂,…,β_m∈R^1×200；

QS4. set miningRadius of sample r 4, radial spacing 2, number of sampling neighborhoods p 8 and sliding window size ω for median filtering_rCount c for 5 × 5₁,c₂,…,c_mAnd s₁,s₂,…,s_mThe robust extended local binary pattern 3-dimensional joint histogram which adopts median filtering and has rotation invariant characteristic and uniform characteristic and is formed by combining 3 descriptors of descriptor based on central point intensity, descriptor based on neighborhood point intensity and descriptor based on radial difference

And

QS5. respectively handle

And

transformed into a 1-dimensional histogram, and then normalized to obtain a feature row vector η₁,η₂,…,η_m∈R^1×200And mu₁,μ₂,…,μ_m∈R^1×200；

Qs6. set the sampling radius r 6, the radial spacing 2, the number of sampling neighbours p 8, and the sliding window size ω for median filtering_rC is counted as 7 × 7₁,c₂,…,c_mAnd s₁,s₂,…,s_mThe robust extended local binary pattern 3-dimensional joint histogram which adopts median filtering and has rotation invariant characteristic and uniform characteristic and is formed by combining 3 descriptors of descriptor based on central point intensity, descriptor based on neighborhood point intensity and descriptor based on radial difference

And

QS7. respectively handle

And

transforming into 1-dimensional histogram, and normalizing them to obtain characteristic row vector theta₁,θ₂,…,θ_m∈R^1×200And

qs8. set the sampling radius r 8, the radial spacing 2, the number of sampling neighborhoods p 8, and the sliding window size ω for median filtering_rCount c for 9 × 9₁,c₂,…,c_mAnd s₁,s₂,…,s_mThe robust extended local binary pattern 3-dimensional joint histogram which adopts median filtering and has rotation invariant characteristic and uniform characteristic and is formed by combining 3 descriptors of descriptor based on central point intensity, descriptor based on neighborhood point intensity and descriptor based on radial difference

And

QS9. eachHandle

And

transformed into a 1-dimensional histogram, and then normalized to obtain a feature row vector k₁,κ₂,…,κ_m∈R^1×200And upsilon₁,υ₂,…,υ_m∈R^1×200；

QS10. construction of c separately₁,c₂,…,c_mAnd s₁,s₂,…,s_mOriginal feature column vector x₁＝[α₁,η₁,θ₁,κ₁]^T,x₂＝[α₂,η₂,θ₂,κ₂]^T,…,x_m＝[α_m,η_m,θ_m,κ_m]^T∈R^800×1And

wherein^TIs a transposition operation;

QS11. constructing an original characteristic matrix X ═ X of the coal rock training sample₁,x₂,…,x_m,x_m+1,x_m+2,…,x_2m]^T∈R^2m×800And its category label matrix Y ∈ R^2m×2Wherein^TFor the transpose operation, the first m rows and columns of data padding of Y are 1, the first m rows and columns of data padding of Y are 0, the last m rows and columns of data padding of Y are 0, and the last m rows and columns of data padding of Y are 1;

QS12, setting a regularization parameter lambda, an iteration number K and an optimal feature number d, and performing regression analysis on X and Y to obtain the advantageD columns j about X in discriminating coal petrography₁,j₂,…,j_dWherein j is not less than 1₁,j₂,…,j_d≤800；

QS13. extract the j' th of X in sequence₁,j₂,…,j_dThe columns are arranged into a final characteristic matrix of the coal rock training sample according to the columns

Wherein X (col. j)₁),X(col·j₂),…,X(col·j_d) J' th of X respectively₁,j₂,…,j_dColumns;

QS14, in the stage of coal rock identification, collecting an unknown class sample image, intercepting a sub-image without a non-coal rock background and carrying out gray processing on the sub-image, and marking the processed unknown class sub-image as q;

qs15. set the sampling radius r 2, the radial spacing 2, the number of sampling neighborhoods p 8, and the sliding window size ω for median filtering_r3 × 3, the robust extended local binary pattern 3-dimensional joint histogram H with the combination of 3 descriptors, namely a descriptor based on central point intensity, a descriptor based on neighborhood point intensity and a descriptor based on radial difference, which adopts median filtering and has rotation invariant characteristic and uniform characteristic_q；

QS16. H_qTransforming into 1-dimensional histogram, and normalizing to obtain characteristic row vector_q∈R^1×200；

QS17. set the sampling radius r 4, radial spacing 2, sampling neighborhood number p 8 and sliding window size ω for median filtering_r5 × 5, the robust extended local binary pattern 3-dimensional joint histogram Z with the rotation invariant characteristic and the uniform characteristic and adopting the median filtering is combined by 3 descriptors of a descriptor based on the central point intensity, a descriptor based on the neighborhood point intensity and a descriptor based on the radial difference_q；

QS18. a handle Z_qThe shape is changed into a 1-dimensional histogram, and then the histogram is normalized to obtain a characteristic row vector rho_q∈R^1×200；

Qs19. set the sampling radius r 6, the radial spacing 2, the number of sampling neighbours p 8, and the sliding window size ω for median filtering_r7 × 7, the robust extended local binary pattern 3-dimensional joint histogram F with the combination of 3 descriptors, namely a descriptor based on central point intensity, a descriptor based on neighborhood point intensity and a descriptor based on radial difference, which adopts median filtering and has rotation invariant characteristic and uniform characteristic, of statistics q_q；

QS20. a handle F_qThe shape is changed into a 1-dimensional histogram, and then the histogram is normalized to obtain a characteristic row vector sigma_q∈R^1×200；

Qs21. set the sampling radius r 8, the radial spacing 2, the number of sampling neighborhoods p 8, and the sliding window size ω for median filtering_r9 × 9, the robust extended local binary pattern 3-dimensional joint histogram T with rotation invariant and uniform characteristics and adopting median filtering and with the rotation invariant characteristic and the uniform characteristic is formed by combining 3 descriptors, namely a descriptor based on central point intensity, a descriptor based on neighborhood point intensity and a descriptor based on radial difference_q；

QS22. respectively handle T_qTransformed into a 1-dimensional histogram, which is then normalized to obtain a feature row vector ξ_q∈R^{1 ×200}；

QS23. construct the raw feature row vector x of q_q＝[_q,ρ_q,σ_q,ξ_q]∈R^1×800；

QS24. successively extract x_qJ (d) of₁,j₂,…,j_dElements, then arranging the final feature row vector constituting q in order

Wherein x_q(j₁),x_q(j₂),…,x_q(j_d) Are respectively x_qJ (d) of₁,j₂,…,j_dAn element;

QS25. by calculating

Judging the coal rock class of q, if the coal rock class satisfies q

Judging q is coal; otherwise, q is determined to be rock, where i is

I is 1,2, …,2m, j is

Column symbol of

The element numbers of (1), (2), (…), (d),

is composed of

Of the elements located in row i and column j,

is composed of

The jth element of (1).

2. The coal-rock identification method based on the extended local binary pattern and the regression analysis according to claim 1, wherein the regression analysis comprises the following steps:

QS1201. construct a symbol matrix U ∈ R^2m×2Then initializing m elements positioned in the first m rows and the 1 st column in the U by using 1, initializing m elements positioned in the first m rows and the 2 nd column in the U by using-1, initializing m elements positioned in the second m rows and the 1 st column in the U by using-1, and initializing m elements positioned in the second m rows and the 2 nd column in the U by using 1;

QS1202. constructing a data matrix X⁰∈R^2m×800Then by

Column by column initialization X⁰Wherein τ is X⁰And X, τ ═ 1,2, …,800,

is X⁰Column τ of, x_τColumn τ of X, mean (X)_τ) Is x_τMean of all elements in;

QS1203, respectively constructing weight matrix W_old,W_new∈R^800×2And an offset vector b_old,b_new∈R^2×1；

Qs1204. first pass W_old＝(X⁰)^T{[X⁰(X⁰)^T+λI_2m]^-1Y initialization W_oldThen through W_new＝W_oldInitializing W_newWherein^TAnd^-1respectively transpose operation and inversion operation, I_2mIs a 2 m-order identity matrix, and lambda is a regularization parameter;

QS1205. construct the data matrix L_b∈R^2×2mThen by

Initialization L_b；

QS1206. first pass

Initialization b_oldThen through b_new＝b_oldInitialization b_newWherein L is_b(col. n) is L_bN is L_bWhere n is 1,2, …,2 m;

QS1207. respectively constructing data matrixes A, B, G ∈ R^2m×2Then initializing them to a zero matrix;

QS1208. construct data matrix X¹∈R^2m×801Then initialize X with X¹First 800 columns of (A), X¹All elements of column 801 are initialized to 1000, i.e., X¹＝[X,1000e_2m]∈R^2m×801Wherein e is_2m∈R^2m×1A column vector with all elements 1;

QS1209. define iteration number k₁And is initialized to 0;

QS1210. by

Update A wherein^TFor transposition operations, e_2m∈R^2m×1A column vector with all elements 1;

QS1211 updating B by replacing the negative element in B with 0, wherein ═ U is A for Hadamard product operation of the matrix;

QS1212, updating G by G ═ B) + Y, wherein &isHadamard product operation of the matrix;

QS1213. describe mathematics as

The optimization problem (W) is recorded as KeyProblem, and the solution W of the KeyProblem is obtained₀∈R^801×2Wherein | · | purple light_2,1Is a matrix of_2,1-a norm;

QS1214. with W₀Update W on the first 800 rows_new；

QS1215. through b_new＝1000×[W₀(row·801)]^TUpdate b_newWherein^TFor transposition operations, W₀(row 801) is W₀Line 801;

QS1216. iteration number k₁Self-increment by 1;

QS1217. if satisfied at the same time

And k₁If < K, then W is used_newUpdate W of the value of_oldBy b_newUpdate of value b_oldThen, a step QS 1210-QS 1217 is performed, in which | · | |. a_FAnd | · | non-conducting phosphor₂The Frobenius norm of the matrix and the 2-norm of the vector are respectively; otherwise, executing step QS 1218;

QS1218. construct a weight square matrix W₂∈R^800×2Then through W₂＝W_new⊙W_newInitializing W₂Wherein ⊙ is a Hadamard product operation of the matrix;

QS1219. construction of weight vector

Then pass through

Updating

Wherein W₂(col.1) and W₂(col. 2) are each W₂Column 1 and column 2;

QS1220. finish regression analysis, return

The sequence number j of the first d elements in the sequence from big to small₁,j₂,…,j_dAs d columns for X.

3. The coal-rock identification method based on the extended local binary pattern and the regression analysis as claimed in claim 2, wherein the solution of the KeyProblem comprises the following steps:

QS121301. construction of data matrix X₊∈R^2m×(801+2m)Then with X¹Initialization of X₊First 801 columns of (1), with 2m order identity matrix I_2mInitialization of X₊The last 2m column of (i.e. X)₊＝[X¹,I_2m]∈R^2m×(801+2m)In which I_2mIs an identity matrix of 2m order;

QS121302. construction of diagonal element vector φ ∈ R^(801+2m)×1Then all elements of phi are initialized to 1;

QS121303. define iteration number k₂And initialized to 0, define and variable Sum and initialized to-1000.000;

QS121304. construction of diagonal matrix Λ∈ R^{(801+2m)×(801+2m)}；

Qs121305. through Λ_ll＝φ_l1,2, …, (801+2m) sequentially updates the main diagonal element of Λ, where l is the element number of phi and the main diagonal element number of Λ, phi_lIs the l-th element of phi, Λ_llThe first major diagonal element of Λ, i.e., Λ_llΛ for the element in row l, column l;

QS121306. separately construct a data matrix W₀∈R^801×2And W₃,W₄∈R^(801+2m)×2Then initializing them to a zero matrix;

QS121307. by

Updating W₃Wherein^TAnd^-1respectively performing transposition operation and inversion operation;

QS121308. with W₄＝W₃⊙W₃Update W of the calculation result of (2)₄Wherein ⊙ is a Hadamard product operation of the matrix;

qs121309. first pass phi ═ W₄(col·1)+W₄(col.2) updating phi, if zero element exists in phi, replacing zero element with 0.000001, wherein W₄(col.1) and W₄(col. 2) are each W₄Column 1 and column 2;

QS121310. iteration number k₂Self-increment by 1;

QS121311. if satisfied simultaneously

And k₂If < K, then pass

Update Sum, then execute stepQS 121305-QS 121311, where l is the element number of φ_lThe l-th element of phi; otherwise, step QS121312 is executed;

QS121312. with W₃The first 801 rows of update W₀；

QS121313. complete the solution of KeyProblem, return the solution W of KeyProblem₀。

Images