CN107239783A - Coal-rock identification method based on extension local binary patterns and regression analysis - Google Patents

Abstract

The invention discloses the Coal-rock identification method based on extension local binary patterns and regression analysis.In the sample training stage, shot by camera, subgraph interception, a series of processing such as gray processing obtain several coal sample subgraphs and rock specimens subgraph, extract the extension local binary patterns characteristic vector of each subgraph, regression analysis is carried out to these characteristic vectors and its class label, obtain feature preferred compositions, feature based preferred compositions, obtain training sample preferred feature vector；In coal petrography cognitive phase, unknown classification sample subgraph is obtained by a series of processing, the extension local binary patterns characteristic vector of unknown classification sample subgraph is extracted, the feature preferred compositions based on the sample training stage obtain unknown classification sample preferred feature vector.Pass through the coal petrography classification belonging to the unknown classification sample of preferred feature vector determination between relatively more unknown classification sample and training sample.The present invention is with the obvious advantage in terms of flexibility, reliability, real-time.

Description

Coal-rock identification method based on extension local binary patterns and regression analysis

Technical field

The present invention relates to the Coal-rock identification method based on extension local binary patterns and regression analysis, belong to image recognition skill Art field.

Background technology

Coal petrography identification refers to by various technological means automatic discrimination coals and rock.In exploitation of coal resources and transported Cheng Zhong, there are many production links is needed to differentiate differentiation coal and rock, and such as coal mining machine roller is highly adjusted, mining mistake Cash etc. is selected in process control, Raw Coal.Since 1950s, the world such as South Africa, Australia, Germany, the U.S., China Main producing coal country expands a series of researchs to Coal-rock identification method, some representational achievements in research is generated in succession, such as Natural Gamma ray probe method, radar detection system, infrared detecting method, active power detection method, vibration signal detection method, voice signal Detection method etc..But there is following common problem in these methods：(1) the various sensings of installation and deployment on existing are needed Device, associated apparatus structure is complicated, and manufacturing cost is high；(2) plant equipment such as coal-winning machine, development machine stress in coal production process Complicated, vibration is violent, serious wear, and sensor deployment is relatively difficult, and its electronic circuit is also easily damaged, and device can It is poor by property；(3) different types of mechanical carrier equipment is directed to, the selection of the type selecting and installation site of sensor has larger area Not, this is accomplished by carrying out personalized customization, therefore its universality is not good.

By the observation to block coal, rock specimens, find coal and rock in terms of color, gloss, texture There is larger difference.When being imaged by existing digital camera to coal and rock, the vision letter of coal and rock Breath will necessarily be just hidden in the digital picture collected, therefore is proposed by excavating the visual information in coal petrography digital picture To distinguish coal and rock.The existing Coal-rock identification method based on image procossing also exists in terms of robustness, discrimination Larger room for promotion.

The content of the invention

In order to overcome the shortcomings of that existing Coal-rock identification method is present, the present invention is proposed based on extension local binary patterns and returned Return the Coal-rock identification method of analysis, this method has the advantages that real-time, discrimination is high, robustness is good, is favorably improved existing For the production efficiency and safe coefficient in colliery.

Coal-rock identification method of the present invention adopts the following technical scheme that realization, including sample training stage and coal petrography are known In the other stage, comprise the following steps that：

RS1. in the sample training stage, collection m width coal sample images and m width rock specimens images, interception are free of non-coal The subgraph of rock background simultaneously carries out gray processing processing to them, and coal sample subgraph and rock specimens subgraph after processing are designated as respectively c₁,c₂,…,c_mAnd s₁,s₂,…,s_m；

RS2. setting sample radius r=2, spaced radial δ=2, sampling Neighborhood Number p=8 and the slip for medium filtering Window size ω_r=3 × 3, c is counted respectively₁,c₂,…,c_mAnd s₁,s₂,…,s_mUse medium filtering and with rotation not Become characteristic and uniform properties by description based on central point intensity, description based on neighborhood point intensity and based on radial difference The robusts extension local binary patterns 3-dimensional joint histograms of this 3 kinds description sub-portfolios of description divided With

RS3. respectivelyWithShape is changed into 1 dimension histogram, and then they are carried out Normalized, obtains feature row vector α₁,α₂,…,α_m∈R^1×200And β₁,β₂,…,β_m∈R^1×200；

RS4. setting sample radius r=4, spaced radial δ=2, sampling Neighborhood Number p=8 and the slip for medium filtering Window size ω_r=5 × 5, c is counted respectively₁,c₂,…,c_mAnd s₁,s₂,…,s_mUse medium filtering and with rotation not Become characteristic and uniform properties by description based on central point intensity, description based on neighborhood point intensity and based on radial difference The robusts extension local binary patterns 3-dimensional joint histograms of this 3 kinds description sub-portfolios of description divided With

RS5. respectivelyWithShape is changed into 1 dimension histogram, and then they are returned One change is handled, and obtains feature row vector η₁,η₂,…,η_m∈R^1×200And μ₁,μ₂,…,μ_m∈R^1×200；

RS6. setting sample radius r=6, spaced radial δ=2, sampling Neighborhood Number p=8 and the slip for medium filtering Window size ω_r=7 × 7, c is counted respectively₁,c₂,…,c_mAnd s₁,s₂,…,s_mUse medium filtering and with rotation not Become characteristic and uniform properties by description based on central point intensity, description based on neighborhood point intensity and based on radial difference The robusts extension local binary patterns 3-dimensional joint histograms of this 3 kinds description sub-portfolios of description divided With

RS7. respectivelyWithShape is changed into 1 dimension histogram, then carries out normalizing to them Change is handled, and obtains feature row vector θ₁,θ₂,…,θ_m∈R^1×200With

RS8. setting sample radius r=8, spaced radial δ=2, sampling Neighborhood Number p=8 and the slip for medium filtering Window size ω_r=9 × 9, c is counted respectively₁,c₂,…,c_mAnd s₁,s₂,…,s_mUse medium filtering and with rotation not Become characteristic and uniform properties by description based on central point intensity, description based on neighborhood point intensity and based on radial difference The robusts extension local binary patterns 3-dimensional joint histograms of this 3 kinds description sub-portfolios of description divided With

RS9. respectivelyWithShape is changed into 1 dimension histogram, and then they are returned One change is handled, and obtains feature row vector κ₁,κ₂,…,κ_m∈R^1×200And υ₁,υ₂,…,υ_m∈R^1×200；

RS10. c is built respectively₁,c₂,…,c_mAnd s₁,s₂,…,s_mPrimitive character 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^800×1With ...,Wherein T is transposition computing；

RS11. the primitive character matrix X=[x of coal petrography training sample are built₁,x₂,…,x_m,x_m+1,x_m+2,…,x_2m]^T∈R^{2m ×800}And its class label matrix Y ∈ R^2m×2, wherein T is transposition computing, and the Y preceding column data of m rows the 1st is filled with 1, Y preceding m rows The rear column data of m rows the 1st that 2nd column data is filled with 0, Y is filled with 0, the Y rear column data of m rows the 2nd and is filled with 1；

RS12., regularization parameter λ, iterations K and preferred feature number d are set, regression analysis is carried out to X and Y, had Beneficial to the d row mark j on X for differentiating coal petrography₁,j₂,…,j_d, wherein 1≤j₁,j₂,…,j_d≤800；

RS13. X jth is successively extracted₁,j₂,…,j_dRow, are then arranged to make up the final feature of coal petrography training sample by row MatrixWherein X (colj₁),X(col·j₂),…,X (col·j_d) be respectively X jth₁,j₂,…,j_dRow；

RS14. in coal petrography cognitive phase, unknown classification sample image is gathered, subgraph of the interception without non-coal petrography background is simultaneously right It carries out gray processing processing, and the unknown classification subgraph after processing is designated as q；

RS15. setting sample radius r=2, spaced radial δ=2, sampling Neighborhood Number p=8 and the slip for medium filtering Window size ω_r=3 × 3, count q use medium filtering and with invariable rotary characteristic and uniform properties by based in Description of heart point intensity, this 3 kinds description sub-portfolios of description based on neighborhood point intensity and description based on radial direction difference Robust extension local binary patterns 3-dimensional joint histogram H_q；

RS16. H_qShape is changed into 1 dimension histogram, and then it is normalized, and obtains feature row vector ε_q∈R^{1 ×200}；

RS17. setting sample radius r=4, spaced radial δ=2, sampling Neighborhood Number p=8 and the slip for medium filtering Window size ω_r=5 × 5, count q use medium filtering and with invariable rotary characteristic and uniform properties by based in Description of heart point intensity, this 3 kinds description sub-portfolios of description based on neighborhood point intensity and description based on radial direction difference Robust extension local binary patterns 3-dimensional joint histogram Z_q；

RS18. Z_qShape is changed into 1 dimension histogram, and then it is normalized, and obtains feature row vector ρ_q∈R^{1 ×200}；

RS19. setting sample radius r=6, spaced radial δ=2, sampling Neighborhood Number p=8 and the slip for medium filtering Window size ω_r=7 × 7, count q use medium filtering and with invariable rotary characteristic and uniform properties by based in Description of heart point intensity, this 3 kinds description sub-portfolios of description based on neighborhood point intensity and description based on radial direction difference Robust extension local binary patterns 3-dimensional joint histogram F_q；

RS20. F_qShape is changed into 1 dimension histogram, and then it is normalized, and obtains feature row vector σ_q∈R^{1 ×200}；

RS21. setting sample radius r=8, spaced radial δ=2, sampling Neighborhood Number p=8 and the slip for medium filtering Window size ω_r=9 × 9, count q use medium filtering and with invariable rotary characteristic and uniform properties by based in Description of heart point intensity, this 3 kinds description sub-portfolios of description based on neighborhood point intensity and description based on radial direction difference Robust extension local binary patterns 3-dimensional joint histogram T_q；

RS22. respectively T_qShape is changed into 1 dimension histogram, and then it is normalized, and obtains feature row vector ξ_q∈ R^1×200；

RS23. q primitive character row vector x is built_q=[ε_q,ρ_q,σ_q,ξ_q]∈R^1×800；

RS24. x is successively extracted_qJth₁,j₂,…,j_dIndividual element, be then arranged to make up in order q final feature row to AmountWherein x_q(j₁),x_q(j₂),…,x_q(j_d) it is respectively x_qJth₁, j₂,…,j_dIndividual element；

RS25. by calculatingQ coal petrography classification is judged, if metSo judge q as coal；Otherwise, it is determined that q is rock, wherein i isRower, i=1,2 ..., 2m, j is Row mark andElement numbers, j=1,2 ..., d,In be located at the i-th row jth arrange element,For J-th of element.

Regression analysis described in step RS12 comprises the following steps：

RS1201. sign matrix U ∈ R are built^2m×2, then it is located at the m element that preceding m rows the 1st are arranged with 1 initialization U, It is located at the m element that preceding m rows the 2nd are arranged with -1 initialization U, is located at the m element that rear m rows the 1st are arranged with -1 initialization U, with 1 Initialize and be located at the m element that rear m rows the 2nd are arranged in U；

RS1202. data matrix X is built⁰∈R^2m×800, then pass throughX is initialized by column⁰, its Middle τ is X⁰With X row mark, τ=1,2 ..., 800,For X⁰τ row, x_τArranged for X τ, mean (x_τ) it is x_τIn all members The average of element；

RS1203. weight matrix W is built respectively_old,W_new∈R^800×2With bias vector b_old,b_new∈R^2×1；

RS1204. W is first passed through_old=(X⁰)^T{[X⁰(X⁰)^T+λI_2m]^-1Y } initialization W_old, then pass through W_new=W_oldInitialization W_new, wherein T and -1 is respectively transposition computing and inversion operation, I_2mFor 2m rank unit matrixs；

RS1205. data matrix L is built_b∈R^2×2m, then pass throughInitialize L_b；

RS1206. first pass throughInitialize b_old, then pass through b_new=b_oldInitialize b_new, wherein L_b(colj) it is L_bJth row, j is L_bRow mark, j=1,2 ..., 2m；

RS1207. data matrix A, B, G ∈ R are built respectively^2m×2, they are then initialized as null matrix；

RS1208. data matrix X is built¹∈R^2m×801, then initialize X with X¹It is preceding 800 row, X¹The 801st row institute There is element to be initialized as 1000, i.e. X¹=[X, 1000e_2m]∈R^2m×801, wherein e_2m∈R^2m×1It is 1 row for all elements Vector；

RS1209. iteration sequence number k is defined₁And it is initialized as 0；

RS1210. pass throughA is updated, wherein T is transposition computing, e_2m∈R^2m×1For all members Element is 1 column vector；

RS1211. B=U ⊙ A are first passed through and update B, then negative value element in B is replaced with 0, wherein ⊙ is matrix Hadamard accumulates computing；

RS1212. G is updated by G=(U ⊙ B)+Y, wherein ⊙ is Hadamard matrix nature computing；

RS1213. it is mathematical description

Optimization problem be designated as KeyProblem, seek KeyProblem solution W₀∈R^801×2, wherein | | | |_2,1For matrix L_2,1- norm, with matrix Ψ ∈ R^u0×v0Exemplified by,I0 be Ψ rower, i0=1, 2 ..., u0, j0 are Ψ row mark, j0=1,2 ..., v0, and Ψ (i0, j0) is to be located at the element that the i-th 0 row jth 0 are arranged in Ψ, u0 and V0 is respectively Ψ line number and columns；

RS1214. W is used₀Preceding 800 row update W_new；

RS1215. b is passed through_new=1000 × [W₀(row·801)]^TUpdate b_new, wherein T is transposition computing, W₀(row· 801) it is W₀The 801st row；

RS1216. iteration sequence number k₁From increasing 1；

If RS1217. meet simultaneouslyAnd k₁<K, then use W_newValue more New W_old, use b_newValue update b_old, step RS1210-RS1217 are then performed, wherein | | | |_FWith | | | |₂Respectively square The Frobenius norms of battle array and 2-norm of vector；Otherwise, step RS1218 is performed；

RS1218. weight square matrices W is built₂∈R^800×2, then pass through W₂=W_new⊙W_newInitialize W₂, wherein ⊙ is Hadamard matrix nature computing；

RS1219. weight vectors are builtThen pass throughUpdateIts Middle W₂And W (col1)₂(col2) it is respectively W₂The 1st row and the 2nd row；

RS1220. regression analysis is completed, is returnedIn by descending order arrange preceding d element sequence number j₁, j₂,…,j_dIt is used as the d row mark on X.

KeyProblem solution comprises the following steps described in step RS1213：

RS121301. data matrix X is built₊∈R^2m×(801+2m), then use X¹Initialize X₊Preceding 801 row, it is single with 2m ranks Bit matrix I_2mInitialize X₊Rear 2m row, i.e. X₊=[X¹,I_2m]∈R^2m×(801+2m), wherein I_2mFor 2m rank unit matrixs；

RS121302. diagonal entry vector φ ∈ R are built^(801+2m)×1, then φ all elements are initialized as 1；

RS121303. iteration sequence number k is defined₂And it is initialized as 0, definition and variable Sum are simultaneously initialized as -1000.000；

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

RS121305. Λ is passed through_jj=φ_j, j=1,2 ..., (801+2m) updates Λ the elements in a main diagonal successively, wherein J is φ element numbers and Λ the elements in a main diagonal sequence number, φ_jFor φ j-th of element, Λ_jjIt is diagonal for Λ j-th of master Line element, that is, Λ_jjTo be located at the element that jth row jth is arranged in Λ；

RS121306. data matrix W is built respectively₀∈R^801×2And W₃,W₄∈R^(801+2m)×2, then they are initialized as Null matrix；

RS121307. pass throughUpdate W₃, wherein T and -1 be respectively transposition computing and Inversion operation；

RS121308. W is used₄=W₃⊙W₃Result of calculation update W₄, wherein ⊙ is Hadamard matrix nature computing；

RS121309. φ=W is first passed through₄(col·1)+W₄(col2) φ is updated, if there is neutral element in φ, is used 0.000001 replaces neutral element, wherein W₄And W (col1)₄(col2) it is respectively W₄The 1st row and the 2nd row；

RS121310. iteration sequence number k₂From increasing 1；

If RS121311. meet simultaneouslyAnd k₂<K, then pass throughUpdate Sum, it is φ element numbers, φ then to perform step RS121305-RS121311, wherein j_jFor φ j-th of element；It is no Then, step RS121312 is performed；

RS121312. W is used₃Preceding 801 row update W₀；

RS121313. KeyProblem solution is completed, KeyProblem solution W is returned₀。

Brief description of the drawings

Fig. 1 is the basic flow sheet of the Coal-rock identification method based on extension local binary patterns and regression analysis；

Fig. 2 is the basic flow sheet of regression analysis of the present invention；

Fig. 3 is the solution W of the present invention for seeking KeyProblem₀Basic flow sheet；

Embodiment

On the basis of experimental analysis is carried out to the image of the ground Coal Gasification of Main Coal Species such as China Henan, Shanxi, Shaanxi and rock kind, this Invention proposes the Coal-rock identification method based on extension local binary patterns and regression analysis, and this method can effectively differentiate coal And rock.

The present invention is described in further detail with reference to the accompanying drawings and detailed description.

Reference picture 1, based on comprising the following steps that for the Coal-rock identification method for extending local binary patterns and regression analysis：

SS1. in the sample training stage, collection m width coal sample images and m width rock specimens images, interception are free of non-coal The subgraph of rock background simultaneously carries out gray processing processing to them, and coal sample subgraph and rock specimens subgraph after processing are designated as respectively c₁,c₂,…,c_mAnd s₁,s₂,…,s_m；

SS2. setting sample radius r=2, spaced radial δ=2, sampling Neighborhood Number p=8 and the slip for medium filtering Window size ω_r=3 × 3, c is counted respectively₁,c₂,…,c_mAnd s₁,s₂,…,s_mUse medium filtering and with rotation not Become characteristic and uniform properties by description based on central point intensity, description based on neighborhood point intensity and based on radial difference The robusts extension local binary patterns 3-dimensional joint histograms of this 3 kinds description sub-portfolios of description divided With

SS3. respectivelyWithShape is changed into 1 dimension histogram, and then they are carried out Normalized, obtains feature row vector α₁,α₂,…,α_m∈R^1×200And β₁,β₂,…,β_m∈R^1×200；

SS4. setting sample radius r=4, spaced radial δ=2, sampling Neighborhood Number p=8 and the slip for medium filtering Window size ω_r=5 × 5, c is counted respectively₁,c₂,…,c_mAnd s₁,s₂,…,s_mUse medium filtering and with rotation not Become characteristic and uniform properties by description based on central point intensity, description based on neighborhood point intensity and based on radial difference The robusts extension local binary patterns 3-dimensional joint histograms of this 3 kinds description sub-portfolios of description divided With

SS5. respectivelyWithShape is changed into 1 dimension histogram, and then they are returned One change is handled, and obtains feature row vector η₁,η₂,…,η_m∈R^1×200And μ₁,μ₂,…,μ_m∈R^1×200；

SS6. setting sample radius r=6, spaced radial δ=2, sampling Neighborhood Number p=8 and the slip for medium filtering Window size ω_r=7 × 7, c is counted respectively₁,c₂,…,c_mAnd s₁,s₂,…,s_mUse medium filtering and with rotation not Become characteristic and uniform properties by description based on central point intensity, description based on neighborhood point intensity and based on radial difference The robusts extension local binary patterns 3-dimensional joint histograms of this 3 kinds description sub-portfolios of description divided With

SS7. respectivelyWithShape is changed into 1 dimension histogram, then carries out normalizing to them Change is handled, and obtains feature row vector θ₁,θ₂,…,θ_m∈R^1×200With

SS8. setting sample radius r=8, spaced radial δ=2, sampling Neighborhood Number p=8 and the slip for medium filtering Window size ω_r=9 × 9, c is counted respectively₁,c₂,…,c_mAnd s₁,s₂,…,s_mUse medium filtering and with rotation not Become characteristic and uniform properties by description based on central point intensity, description based on neighborhood point intensity and based on radial difference The robusts extension local binary patterns 3-dimensional joint histograms of this 3 kinds description sub-portfolios of description divided With

SS9. respectivelyWithShape is changed into 1 dimension histogram, and then they are returned One change is handled, and obtains feature row vector κ₁,κ₂,…,κ_m∈R^1×200And υ₁,υ₂,…,υ_m∈R^1×200；

SS10. c is built respectively₁,c₂,…,c_mAnd s₁,s₂,…,s_mPrimitive character 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^800×1With ...,Wherein T is transposition computing；

SS11. the primitive character matrix X=[x of coal petrography training sample are built₁,x₂,…,x_m,x_m+1,x_m+2,…,x_2m]^T∈R^{2m ×800}And its class label matrix Y ∈ R^2m×2, wherein T is transposition computing, and the Y preceding column data of m rows the 1st is filled with 1, Y preceding m rows The rear column data of m rows the 1st that 2nd column data is filled with 0, Y is filled with 0, the Y rear column data of m rows the 2nd and is filled with 1；

SS12., regularization parameter λ, iterations K and preferred feature number d are set, regression analysis is carried out to X and Y, had Beneficial to the d row mark j on X for differentiating coal petrography₁,j₂,…,j_d, wherein 1≤j₁,j₂,…,j_d≤800；

SS13. X jth is successively extracted₁,j₂,…,j_dRow, are then arranged to make up the final feature of coal petrography training sample by row MatrixX(col·j₂),…,X(col·j_d)]∈R^2m×d, wherein X (colj₁),X(col· j₂),…,X(col·j_d) be respectively X jth₁,j₂,…,j_dRow；

SS14. in coal petrography cognitive phase, unknown classification sample image is gathered, subgraph of the interception without non-coal petrography background is simultaneously right It carries out gray processing processing, and the unknown classification subgraph after processing is designated as q；

SS15. setting sample radius r=2, spaced radial δ=2, sampling Neighborhood Number p=8 and the slip for medium filtering Window size ω_r=3 × 3, count q use medium filtering and with invariable rotary characteristic and uniform properties by based in Description of heart point intensity, this 3 kinds description sub-portfolios of description based on neighborhood point intensity and description based on radial direction difference Robust extension local binary patterns 3-dimensional joint histogram H_q；

SS16. H_qShape is changed into 1 dimension histogram, and then it is normalized, and obtains feature row vector ε_q∈R^{1 ×200}；

SS17. setting sample radius r=4, spaced radial δ=2, sampling Neighborhood Number p=8 and the slip for medium filtering Window size ω_r=5 × 5, count q use medium filtering and with invariable rotary characteristic and uniform properties by based in Description of heart point intensity, this 3 kinds description sub-portfolios of description based on neighborhood point intensity and description based on radial direction difference Robust extension local binary patterns 3-dimensional joint histogram Z_q；

SS18. Z_qShape is changed into 1 dimension histogram, and then it is normalized, and obtains feature row vector ρ_q∈R^{1 ×200}；

SS19. setting sample radius r=6, spaced radial δ=2, sampling Neighborhood Number p=8 and the slip for medium filtering Window size ω_r=7 × 7, count q use medium filtering and with invariable rotary characteristic and uniform properties by based in Description of heart point intensity, this 3 kinds description sub-portfolios of description based on neighborhood point intensity and description based on radial direction difference Robust extension local binary patterns 3-dimensional joint histogram F_q；

SS20. F_qShape is changed into 1 dimension histogram, and then it is normalized, and obtains feature row vector σ_q∈R^{1 ×200}；

SS21. setting sample radius r=8, spaced radial δ=2, sampling Neighborhood Number p=8 and the slip for medium filtering Window size ω_r=9 × 9, count q use medium filtering and with invariable rotary characteristic and uniform properties by based in Description of heart point intensity, this 3 kinds description sub-portfolios of description based on neighborhood point intensity and description based on radial direction difference Robust extension local binary patterns 3-dimensional joint histogram T_q；

SS22. respectively T_qShape is changed into 1 dimension histogram, and then it is normalized, and obtains feature row vector ξ_q∈ R^1×200；

SS23. q primitive character row vector x is built_q=[ε_q,ρ_q,σ_q,ξ_q]∈R^1×800；

SS24. x is successively extracted_qJth₁,j₂,…,j_dIndividual element, be then arranged to make up in order q final feature row to AmountWherein x_q(j₁),x_q(j₂),…,x_q(j_d) it is respectively x_qJth₁, j₂,…,j_dIndividual element；

SS25. by calculatingQ coal petrography classification is judged, if metSo judge q as coal；Otherwise, it is determined that q is rock, wherein i isRower, i=1,2 ..., 2m, j is Row mark andElement numbers, j=1,2 ..., d,ForIn be located at the i-th row jth arrange element,For J-th of element.

Reference picture 2, regression analysis described in step SS12 is comprised the following steps that：

SS1201. sign matrix U ∈ R are built^2m×2, then it is located at the m element that preceding m rows the 1st are arranged with 1 initialization U, It is located at the m element that preceding m rows the 2nd are arranged with -1 initialization U, is located at the m element that rear m rows the 1st are arranged with -1 initialization U, with 1 Initialize and be located at the m element that rear m rows the 2nd are arranged in U；

SS1202. data matrix X is built⁰∈R^2m×800, then pass throughX is initialized by column⁰, its Middle τ is X⁰With X row mark, τ=1,2 ..., 800,For X⁰τ row, x_τArranged for X τ, mean (x_τ) it is x_τIn all members The average of element；

SS1203. weight matrix W is built respectively_old,W_new∈R^800×2With bias vector b_old,b_new∈R^2×1；

SS1204. W is first passed through_old=(X⁰)^T{[X⁰(X⁰)^T+λI_2m]^-1Y } initialization W_old, then pass through W_new=W_oldInitialization W_new, wherein T and -1 is respectively transposition computing and inversion operation, I_2mFor 2m rank unit matrixs；

SS1205. data matrix L is built_b∈R^2×2m, then pass throughInitialize L_b；

SS1206. first pass throughInitialize b_old, then pass through b_new=b_oldInitialize b_new, wherein L_b(colj) it is L_bJth row, j is L_bRow mark, j=1,2 ..., 2m；

SS1207. data matrix A, B, G ∈ R are built respectively^2m×2, they are then initialized as null matrix；

SS1208. data matrix X is built¹∈R^2m×801, then initialize X with X¹It is preceding 800 row, X¹The 801st row institute There is element to be initialized as 1000, i.e. X¹=[X, 1000e_2m]∈R^2m×801, wherein e_2m∈R^2m×1It is 1 row for all elements Vector；

SS1209. iteration sequence number k is defined₁And it is initialized as 0；

SS1210. pass throughA is updated, wherein T is transposition computing, e_2m∈R^2m×1For all members Element is 1 column vector；

SS1211. B=U ⊙ A are first passed through and update B, then negative value element in B is replaced with 0, wherein ⊙ is matrix Hadamard accumulates computing；

SS1212. G is updated by G=(U ⊙ B)+Y, wherein ⊙ is Hadamard matrix nature computing；

SS1213. it is mathematical description

Optimization problem be designated as KeyProblem, seek KeyProblem solution W₀∈R^801×2, wherein | | | |_2,1For matrix L_2,1- norm, with matrix Ψ ∈ R^u0×v0Exemplified by,I0 be Ψ rower, i0=1, 2 ..., u0, j0 are Ψ row mark, j0=1,2 ..., v0, and Ψ (i0, j0) is to be located at the element that the i-th 0 row jth 0 are arranged in Ψ, u0 and V0 is respectively Ψ line number and columns；

SS1214. W is used₀Preceding 800 row update W_new；

SS1215. b is passed through_new=1000 × [W₀(row·801)]^TUpdate b_new, wherein T is transposition computing, W₀(row· 801) it is W₀The 801st row；

SS1216. iteration sequence number k₁From increasing 1；

If SS1217. meet simultaneouslyAnd k₁<K, then use W_newValue more New W_old, use b_newValue update b_old, step SS1210-SS1217 are then performed, wherein | | | |_FWith | | | |₂Respectively square The Frobenius norms of battle array and 2-norm of vector；Otherwise, step SS1218 is performed；

SS1218. weight square matrices W is built₂∈R^800×2, then pass through W₂=W_new⊙W_newInitialize W₂, wherein ⊙ is Hadamard matrix nature computing；

SS1219. weight vectors are builtThen pass throughUpdateIts Middle W₂And W (col1)₂(col2) it is respectively W₂The 1st row and the 2nd row；

SS1220. regression analysis is completed, is returnedIn by descending order arrange preceding d element sequence number j₁, j₂,…,j_dIt is used as the d row mark on X.

Reference picture 3, seeks the solution W of KeyProblem described in step SS1213₀Comprise the following steps that：

SS121301. data matrix X is built₊∈R^2m×(801+2m), then use X¹Initialize X₊Preceding 801 row, it is single with 2m ranks Bit matrix I_2mInitialize X₊Rear 2m row, i.e. X₊=[X¹,I_2m]∈R^2m×(801+2m), wherein I_2mFor 2m rank unit matrixs；

SS121302. diagonal entry vector φ ∈ R are built^(801+2m)×1, then φ all elements are initialized as 1；

SS121303. iteration sequence number k is defined₂And it is initialized as 0, definition and variable Sum are simultaneously initialized as -1000.000；

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

SS121305. Λ is passed through_jj=φ_j, j=1,2 ..., (801+2m) updates Λ the elements in a main diagonal successively, wherein J is φ element numbers and Λ the elements in a main diagonal sequence number, φ_jFor φ j-th of element, Λ_jjIt is diagonal for Λ j-th of master Line element, that is, Λ_jjTo be located at the element that jth row jth is arranged in Λ；

SS121306. data matrix W is built respectively₀∈R^801×2And W₃,W₄∈R^(801+2m)×2, then they are initialized as Null matrix；

SS121307. pass throughUpdate W₃, wherein T and -1 be respectively transposition computing and Inversion operation；

SS121308. W is used₄=W₃⊙W₃Result of calculation update W₄, wherein ⊙ is Hadamard matrix nature computing；

SS121309. φ=W is first passed through₄(col·1)+W₄(col2) φ is updated, if there is neutral element in φ, is used 0.000001 replaces neutral element, wherein W₄And W (col1)₄(col2) it is respectively W₄The 1st row and the 2nd row；

SS121310. iteration sequence number k₂From increasing 1；

If SS121311. meet simultaneouslyAnd k₂<K, then pass throughUpdate Sum, it is φ element numbers, φ then to perform step SS121305-SS121311, wherein j_jFor φ j-th of element；It is no Then, step SS121312 is performed；

SS121312. W is used₃Preceding 801 row update W₀；

SS121313. KeyProblem solution is completed, KeyProblem solution W is returned₀。

It is pointed out that embodiment described above is used to further illustrate the present invention, embodiment is not construed as Limit the scope of the present invention.

Claims (3)

1. the Coal-rock identification method based on extension local binary patterns and regression analysis, it is characterised in that comprise the following steps：

QS1. in the sample training stage, collection m width coal sample images and m width rock specimens images, interception is without the non-coal petrography back of the body The subgraph of scape simultaneously carries out gray processing processing to them, and coal sample subgraph and rock specimens subgraph after processing are designated as c respectively₁, c₂,…,c_mAnd s₁,s₂,…,s_m；

QS2. sample radius r=2, spaced radial δ=2, sampling Neighborhood Number p=8 and the sliding window for medium filtering are set Size ω_r=3 × 3, c is counted respectively₁,c₂,…,c_mAnd s₁,s₂,…,s_mUse medium filtering and special with invariable rotary Property and uniform properties by based on central point intensity description son, based on neighborhood point intensity description son and based on radial direction difference The robust extension local binary patterns 3-dimensional joint histogram of this 3 kinds description sub-portfolios of descriptionWith

QS3. respectivelyWithShape is changed into 1 dimension histogram, then carries out normalizing to them Change is handled, and obtains feature row vector α₁,α₂,…,α_m∈R^1×200And β₁,β₂,…,β_m∈R^1×200；

QS4. sample radius r=4, spaced radial δ=2, sampling Neighborhood Number p=8 and the sliding window for medium filtering are set Size ω_r=5 × 5, c is counted respectively₁,c₂,…,c_mAnd s₁,s₂,…,s_mUse medium filtering and special with invariable rotary Property and uniform properties by based on central point intensity description son, based on neighborhood point intensity description son and based on radial direction difference The robust extension local binary patterns 3-dimensional joint histogram of this 3 kinds description sub-portfolios of descriptionWith

QS5. respectivelyWithShape is changed into 1 dimension histogram, and then they are normalized Processing, obtains feature row vector η₁,η₂,…,η_m∈R^1×200And μ₁,μ₂,…,μ_m∈R^1×200；

QS6. sample radius r=6, spaced radial δ=2, sampling Neighborhood Number p=8 and the sliding window for medium filtering are set Size ω_r=7 × 7, c is counted respectively₁,c₂,…,c_mAnd s₁,s₂,…,s_mUse medium filtering and special with invariable rotary Property and uniform properties by based on central point intensity description son, based on neighborhood point intensity description son and based on radial direction difference The robust extension local binary patterns 3-dimensional joint histogram of this 3 kinds description sub-portfolios of descriptionWith

QS7. respectivelyWithShape is changed into 1 dimension histogram, and place then is normalized to them Reason, obtains feature row vector θ₁,θ₂,…,θ_m∈R^1×200With

QS8. sample radius r=8, spaced radial δ=2, sampling Neighborhood Number p=8 and the sliding window for medium filtering are set Size ω_r=9 × 9, c is counted respectively₁,c₂,…,c_mAnd s₁,s₂,…,s_mUse medium filtering and special with invariable rotary Property and uniform properties by based on central point intensity description son, based on neighborhood point intensity description son and based on radial direction difference The robust extension local binary patterns 3-dimensional joint histogram of this 3 kinds description sub-portfolios of descriptionWith

QS9. respectivelyWithShape is changed into 1 dimension histogram, and then they are normalized Processing, obtains feature row vector κ₁,κ₂,…,κ_m∈R^1×200And υ₁,υ₂,…,υ_m∈R^1×200；

QS10. c is built respectively₁,c₂,…,c_mAnd s₁,s₂,…,s_mPrimitive character 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^800×1With Wherein^TFor transposition computing；

QS11. the primitive character matrix X=[x of coal petrography training sample are built₁,x₂,…,x_m,x_m+1,x_m+2,…,x_2m]^T∈R^2m×800 And its class label matrix Y ∈ R^2m×2, wherein^TFor transposition computing, the Y preceding column data of m rows the 1st is filled with 1, Y preceding m rows the 2nd The rear column data of m rows the 1st that column data is filled with 0, Y is filled with 0, the Y rear column data of m rows the 2nd and is filled with 1；

QS12., regularization parameter λ, iterations K and preferred feature number d are set, regression analysis is carried out to X and Y, is conducive to Differentiate the d row mark j on X of coal petrography₁,j₂,…,j_d, wherein 1≤j₁,j₂,…,j_d≤800；

QS13. X jth is successively extracted₁,j₂,…,j_dRow, are then lined up the final eigenmatrix of coal petrography training sample by rowWherein X (colj₁),X(col·j₂),…,X (col·j_d) be respectively X jth₁,j₂,…,j_dRow；

QS14. in coal petrography cognitive phase, unknown classification sample image is gathered, subgraph of the interception without non-coal petrography background simultaneously enters to it Row gray processing processing, the unknown classification subgraph after processing is designated as q；

QS15. sample radius r=2, spaced radial δ=2, sampling Neighborhood Number p=8 and the sliding window for medium filtering are set Size ω_r=3 × 3, count q use medium filtering and with invariable rotary characteristic and uniform properties by based on central point Description of intensity, this 3 kinds description sub-portfolios of description based on neighborhood point intensity and description based on radial direction difference are formed Robust extension local binary patterns 3-dimensional joint histogram H_q；

QS16. H_qShape is changed into 1 dimension histogram, and then it is normalized, and obtains feature row vector ε_q∈R^1×200；

QS17. sample radius r=4, spaced radial δ=2, sampling Neighborhood Number p=8 and the sliding window for medium filtering are set Size ω_r=5 × 5, count q use medium filtering and with invariable rotary characteristic and uniform properties by based on central point Description of intensity, this 3 kinds description sub-portfolios of description based on neighborhood point intensity and description based on radial direction difference are formed Robust extension local binary patterns 3-dimensional joint histogram Z_q；

QS18. Z_qShape is changed into 1 dimension histogram, and then it is normalized, and obtains feature row vector ρ_q∈R^1×200；

QS19. sample radius r=6, spaced radial δ=2, sampling Neighborhood Number p=8 and the sliding window for medium filtering are set Size ω_r=7 × 7, count q use medium filtering and with invariable rotary characteristic and uniform properties by based on central point Description of intensity, this 3 kinds description sub-portfolios of description based on neighborhood point intensity and description based on radial direction difference are formed Robust extension local binary patterns 3-dimensional joint histogram F_q；

QS20. F_qShape is changed into 1 dimension histogram, and then it is normalized, and obtains feature row vector σ_q∈R^1×200；

QS21. sample radius r=8, spaced radial δ=2, sampling Neighborhood Number p=8 and the sliding window for medium filtering are set Size ω_r=9 × 9, count q use medium filtering and with invariable rotary characteristic and uniform properties by based on central point Description of intensity, this 3 kinds description sub-portfolios of description based on neighborhood point intensity and description based on radial direction difference are formed Robust extension local binary patterns 3-dimensional joint histogram T_q；

QS22. respectively T_qShape is changed into 1 dimension histogram, and then it is normalized, and obtains feature row vector ξ_q∈R^{1 ×200}；

QS23. q primitive character row vector x is built_q=[ε_q,ρ_q,σ_q,ξ_q]∈R^1×800；

QS24. x is successively extracted_qJth₁, j₂..., j_dIndividual element, is then arranged to make up q final feature row vector in orderWherein x_q(j₁),x_q(j₂),…,x_q(j_d) it is respectively x_qJth₁, j₂,…,j_dIndividual element；

QS25. by calculatingQ coal petrography classification is judged, if meeting Then judge q as coal；Otherwise, it is determined that q is rock, wherein i isRower, i=1,2 ..., 2m, j isRow mark and Element numbers, j=1,2 ..., d,ForIn be located at the i-th row jth arrange element,ForJ-th yuan Element.

2. the Coal-rock identification method according to claim 1 based on extension local binary patterns and regression analysis, its feature It is, the regression analysis comprises the following steps：

QS1201. sign matrix U ∈ R are built^2m×2, then it is located at the m element that preceding m rows the 1st are arranged with 1 initialization U, with the beginning of -1 It is located at the m element that preceding m rows the 2nd are arranged in beginningization U, is located at the m element that rear m rows the 1st are arranged with -1 initialization U, with 1 initialization It is located at the m element that rear m rows the 2nd are arranged in U；

QS1202. data matrix X is built⁰∈R^2m×800, then pass throughX is initialized by column⁰, wherein τ is X⁰With X row mark, τ=1,2 ..., 800,For X⁰τ row, x_τArranged for X τ, mean (x_τ) it is x_τMiddle all elements Average；

QS1203. weight matrix W is built respectively_old,W_new∈R^800×2With bias vector b_old,b_new∈R^2×1；

QS1204. first pass throughInitialize W_old, then pass through W_new=W_oldInitially Change W_new, wherein^TWith^-1Respectively transposition computing and inversion operation, I_2mFor 2m rank unit matrixs；

QS1205. data matrix L is built_b∈R^2×2m, then pass throughInitialize L_b；

QS1206. first pass throughInitialize b_old, then pass through b_new=b_oldInitialize b_new, wherein L_b (colj) it is L_bJth row, j is L_bRow mark, j=1,2 ..., 2m；

QS1207. data matrix A, B, G ∈ R are built respectively^2m×2, they are then initialized as null matrix；

QS1208. data matrix X is built¹∈R^2m×801, then initialize X with X¹It is preceding 800 row, X¹The 801st row all members Element is initialized as 1000, i.e. X¹=[X, 1000e_2m]∈R^2m×801, wherein e_2m∈R^2m×1For all elements be 1 row to Amount；

QS1209. iteration sequence number k is defined₁And it is initialized as 0；

QS1210. pass throughA is updated, wherein^TFor transposition computing, e_2m∈R^2m×1It is for all elements 1 column vector；

QS1211. first pass through B=U ⊙ A update B, then with 0 replace B in negative value element, wherein ⊙ be Hadamard matrix nature Computing；

QS1212. G is updated by G=(U ⊙ B)+Y, wherein ⊙ is Hadamard matrix nature computing；

QS1213. it is mathematical description

Optimization problem be designated as KeyProblem, seek KeyProblem solution W₀∈R^801×2, wherein | | | |_2,1For matrix l_2,1- norm；

QS1214. W is used₀Preceding 800 row update W_new；

QS1215. b is passed through_new=1000 × [W₀(row·801)]^TUpdate b_new, wherein^TFor transposition computing, W₀(row801) it is W₀The 801st row；

QS1216. iteration sequence number k₁From increasing 1；

If QS1217. meet simultaneouslyAnd k₁<K, then use W_newValue update W_old, Use b_newValue update b_old, step QS1210-QS1217 are then performed, wherein | | | |_FWith | | | |₂Respectively matrix 2-norm of Frobenius norms and vector；Otherwise, step QS1218 is performed；

QS1218. weight square matrices W is built₂∈R^800×2, then pass through W₂=W_new⊙W_newInitialize W₂, wherein ⊙ is matrix Hadamard product computing；

QS1219. weight vectors are builtThen pass throughUpdateWherein W₂ And W (col1)₂(col2) it is respectively W₂The 1st row and the 2nd row；

QS1220. regression analysis is completed, is returnedIn by descending order arrange preceding d element sequence number j₁,j₂,…,j_d It is used as the d row mark on X.

3. the Coal-rock identification method according to claim 2 based on extension local binary patterns and regression analysis, its feature It is, the solution of the KeyProblem comprises the following steps：

QS121301. data matrix X is built₊∈R^2m×(801+2m), then use X¹Initialize X₊It is preceding 801 row, with 2m rank unit matrixs I_2mInitialize X₊Rear 2m row, i.e. X₊=[X¹,I_2m]∈R^2m×(801+2m), wherein I_2mFor 2m rank unit matrixs；

QS121302. diagonal entry vector φ ∈ R are built^(801+2m)×1, φ all elements are then initialized as 1；

QS121303. iteration sequence number k is defined₂And it is initialized as 0, definition and variable Sum are simultaneously initialized as -1000.000；

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

QS121305. Λ is passed through_jj=φ_j, j=1,2 ..., (801+2m) updates Λ the elements in a main diagonal successively, and wherein j is φ element numbers and Λ the elements in a main diagonal sequence number, φ_jFor φ j-th of element, Λ_jjFor Λ j-th of leading diagonal Element, that is, Λ_jjTo be located at the element that jth row jth is arranged in Λ；

QS121306. data matrix W is built respectively₀∈R^801×2And W₃,W₄∈R^(801+2m)×2, they are then initialized as zero moment Battle array；

QS121307. pass throughUpdate W₃, wherein^TWith^-1Respectively transposition computing and fortune of inverting Calculate；

QS121308. W is used₄=W₃⊙W₃Result of calculation update W₄, wherein ⊙ is Hadamard matrix nature computing；

QS121309. φ=W is first passed through₄(col·1)+W₄(col2) φ is updated, if there is neutral element in φ, is used 0.000001 replaces neutral element, wherein W₄And W (col1)₄(col2) it is respectively W₄The 1st row and the 2nd row；

QS121310. iteration sequence number k₂From increasing 1；

If QS121311. meet simultaneouslyAnd k₂<K, then pass throughSum is updated, so It is φ element numbers, φ to perform step QS121305-QS121311, wherein j afterwards_jFor φ j-th of element；Otherwise, perform Step QS121312；

QS121312. W is used₃Preceding 801 row update W₀；

QS121313. KeyProblem solution is completed, KeyProblem solution W is returned₀。