CN109523566A - A kind of automatic division method of Sandstone Slice micro-image - Google Patents

A kind of automatic division method of Sandstone Slice micro-image Download PDF

Info

Publication number
CN109523566A
CN109523566A CN201811085809.3A CN201811085809A CN109523566A CN 109523566 A CN109523566 A CN 109523566A CN 201811085809 A CN201811085809 A CN 201811085809A CN 109523566 A CN109523566 A CN 109523566A
Authority
CN
China
Prior art keywords
image
image block
sandstone
follows
type
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201811085809.3A
Other languages
Chinese (zh)
Inventor
姜枫
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Individual
Original Assignee
Individual
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Individual filed Critical Individual
Priority to CN201811085809.3A priority Critical patent/CN109523566A/en
Publication of CN109523566A publication Critical patent/CN109523566A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/136—Segmentation; Edge detection involving thresholding
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/90—Determination of colour characteristics
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10056—Microscopic image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30108—Industrial image inspection
    • G06T2207/30132—Masonry; Concrete

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a kind of automatic division method of Sandstone Slice micro-image, it the steps include: 1) using super-pixel segmentation technology, be image block by sandstone cross-polarized light micro-image pre-segmentation；2) it is based on cross-polarized light micro-image, extracts image block color characteristic and textural characteristics and construction feature vector；3) based on single polarized light microscopy Slice Image, image boundary feature is extracted using bound test technology；4) it is based on sandstone particle specimens data set Training Support Vector Machines classifier；5) trained listening group is used, predicts that each image block belongs to the probability of quartz, feldspar, landwaste, image block type is marked by preset condition；6) type of non-marking types image block is predicted using label propagation algorithm；7) merging type is identical, the lower adjacent image block of boundary characteristic intensity.This method uses image processing techniques, machine learning method and data digging method, the cross-polarized light micro-image and single resonance offset effect obtained in conjunction with same Sandstone Slice, the mineral grain for including in Sandstone Slice is divided automatically, the time manually divided the work and economic cost are reduced, segmentation accuracy is improved.

Description

A kind of automatic division method of Sandstone Slice micro-image
Technical field
The present invention relates to the automatic segmentation problems of Sandstone Slice micro-image, in conjunction with cross-polarized light image and single polarisation figure Picture is realized Sandstone Slice is micro- with image processing techniques and machine learning method by the classification processing to image block Mineral grain included in image is divided into individual region one by one.
Background technique
Sandstone is wide in the distributed pole of nature, is the main reservoir for constituting petroleum, natural gas and underground water.Sandstone Slice Identification can analyze sandstone mineral constituent and mineral content, the type and structure feature of sandstone be obtained, in oil gas reservoir Layer Detection There is important application value with assessment aspect.The basis of Sandstone Slice identification is to carry out image to Sandstone Slice micro-image Segmentation.
Existing Sandstone Slice micro-image segmentation takes time and effort, has to profile and experience based on manually dividing Compared with strong dependency, and its result has nonrepeatability.It include a large amount of mineral grain in Sandstone Slice micro-image, Boundary characteristic between each particle is fuzzy, brings difficulty to image segmentation.In addition, comprising complicated more inside sandstone particle The crystal microstructure of sample is often mistakenly thought of as being granule boundary, also has an impact to the result of image segmentation.
Summary of the invention
Technical problem to be solved by the invention is to provide a kind of automatic division method of Sandstone Slice micro-image, the party Method combination cross-polarized light image and single polarized light image realize that Sandstone Slice is aobvious with image processing techniques and machine learning method The automatic segmentation of micro- image.
In order to achieve the above objectives, this method uses following step:
1) sandstone cross-polarized light microsection image is read, uses super-pixel segmentation technology by its pre-segmentation for image block；
2) it is based on cross-polarized light microsection image, extracts image block color characteristic and textural characteristics；
3) sandstone list polarized light microscopy Slice Image is read, image boundary is extracted using bound test technology；In conjunction with pre-segmentation Image block extracts image block boundaries feature；
4) it is based on sandstone particle specimens data set Training Support Vector Machines classifier；
5) trained listening group is used, the probability that each image block belongs to quartz, feldspar, landwaste is calculated；By default Threshold value determines image block classification, higher than the image block marking types of threshold value, remaining image block not marking types；
6) similar matrix is constructed according to the color characteristic of image block and textural characteristics, the use of label propagation algorithm is step 5) The image block marking types of non-marking types；
7) the adjacent image block that type is identical, boundary characteristic is less than threshold value merges, and forms complete mineral grain.
Above-mentioned steps 1) in by image pre-segmentation be image block detailed process be: first reading sandstone cross-polarized light it is micro- Image, along image x-axis and y-axis both direction withFor interval (wherein number of pixels in N image), image is divided At K square area, each region is considered as an initial clustering；Then according to pixel color feature and coordinate value, iteratively It calculates it and belongs to cluster；After iteration ends, the cluster that each pixel is constituted is an image block.
It is as follows to calculate the method clustered belonging to each pixel: according to K initial clustering, calculating the center of each cluster {ck}k∈[1,K], and with ckCentered on search for the set PS of all pixels around it within the scope of 2H × 2H；To pixel each in PS Point i calculates its distance center ckDistance, update the distance if the distance is less than the distance of last iteration, and being arranged should Cluster belonging to pixel is k；After primary to all cluster centre iteration, each cluster centre and new and old cluster are recalculated The Euclidean distance at center stops algorithm if distance is less than preset threshold, otherwise enters next iteration.
Above-mentioned calculating two pixels i and ckBetween distance method it is as follows:
Wherein rc、gcAnd bcFor ckRGB color characteristic value, xcAnd ycFor ckX-axis and y-axis coordinate；ri、giAnd biFor picture The RGB color characteristic value of vegetarian refreshments i, xiAnd yiFor the x-axis and y-axis coordinate of pixel i, m is preferably 8.
Above-mentioned steps 2) in calculate image block color characteristic detailed process be: image block is converted into grayscale image first Picture calculates the color characteristic hist based on histogram, and the feature based on statistics, comprising: mean value Mean, variance Five Variance, intermediate value Median, mode Mode and mean absolute deviation MAD statistical indicators.
Color characteristic hist calculation method based on histogram is as follows: 0~255 gray scale is divided into 16 rank (each grades Bao Han 16 gray scales), calculate the number of pixels { cnt that each grey level is included in image block1,cnt2,...,cnt16, And the vector is normalized to obtain hist={ hist1,hist2,...,hist16}；
Feature calculation formula based on statistics is as follows, the number of pixels for wherein including in N representative image block, IpFor pixel p Gray value:
Median=mid (I1,I2,…,IN), wherein mid () is median function
By above-mentioned calculating, the color characteristic of each image block is the feature vector of 20 dimensions.
Above-mentioned steps 2) in calculate image block textural characteristics detailed process be: image block is converted into grayscale image first Picture, and calculate gray level co-occurrence matrixes；Textural characteristics are calculated according to gray level co-occurrence matrixes.
The method for calculating image grayscale co-occurrence matrix is as follows: arbitrarily taking a point p1=(x, y) in the picture and deviates its One point p2=(x+a, y+b) calculates point to the gray value (f1, f2) of (p1, p2)；It allows point p1 to move in entire image, obtains Different (f1, f2) values；Count the number of each (f1, f2) value appearance；Matrix P is constructed, P (i, j) is defined as gray value The number that (i, j) occurs；Matrix P is normalized, i.e., total time occurred element each in matrix divided by all gray values Number.Define (a, b) respectively (0,1), (1,0), (1,1) and (- 1,1), therefore obtain 4 gray level co-occurrence matrixes.
According to gray level co-occurrence matrixes calculate textural characteristics method it is as follows: calculate separately ENERGY E nergy, entropy Entropy, Maximum probability MaxProbability, contrast C ontrast, reciprocal difference square IDM index, formula are as follows:
The textural characteristics that each gray matrix obtains include 5 elements, therefore the textural characteristics of each image block are 20 dimensions Feature vector.
Above-mentioned steps 3) in using bound test technology extract image boundary detailed process be: first reading sandstone list it is inclined Light microsection image converts it to gray level image, and pre-processes to image；Then using Canny operator detection figure Boundary as in；Finally, calculating the boundary characteristic of image block according to image boundary.
It is as follows that pretreated method is carried out to image: image being smoothed using Gaussian filter first；Then Gamma correction is carried out to image；Finally carry out histogram equalization.
The method for carrying out gamma correction to image is: for pixel p, enabling its gray value is Ip, first to IpNormalization, Formula are as follows: x=(Ip+0.5)/256；Then x is pre-compensated for, formula are as follows: y=x1/γ, select γ=2.2；Finally carry out Renormalization, formula are as follows: z=y*256-0.5, z are the gray value of pixel p after gamma correction.
The method for carrying out histogram equalization to image is: enabling in image has N number of pixel, and image grayscale is divided into l Gray level, l are preferably 16, and calculating gray level is rkPixel number nk, then the probability that k-th of gray level occurs is P (rk) =nk/ N, k=0,1 ..., l-1, histogram equalization formula are as follows:
Wherein skBe gray level be k pixel through transformation after gray scale.
Above-mentioned steps 4) in the detailed process of Training Support Vector Machines classifier be: firstly, collecting quartz, feldspar, landwaste The image sample data collection of particle；The quantity for counting every class sandstone particle image implements image to the grain type of negligible amounts The method of enhancing is generated additional image pattern by the methods of translation, overturning, rotation, scaling, makes three types particle figure Picture balances between reaching class；For each sandstone particle specimens image, color characteristic and texture are calculated according to the method for step 2) Feature, and constitutive characteristic vector；It is marked according to particle image classification, constructs classifier training collection；Finally, based on image training Collection, Training Support Vector Machines classifier.
Above-mentioned steps 5) in be to the detailed process that image block type is labeled: using the trained support of step 4) The color feature vector and texture feature vector that step 2) is calculated in vector machine classifier predict, obtain probability to Measure p={ p1,p2,p3, it is the probability that the image block belongs to quartz, feldspar and landwaste respectively；Enable pmax=max (p1,p2, p3), As pmax >=0.7, the type for marking image block is the type of sandstone particle corresponding to pmax.
Above-mentioned steps 6) in treatment process that the type for not marking image block is predicted be: first by step 1) To each image block regard a vertex as, construct similar matrix；Then using label propagation algorithm to the figure of non-marking types As block carries out type mark.
The method of building similar matrix is: having the image block of marking types to constitute labeled data step 5), is denoted as (x1, y1),(x2,y2),…,(xL,yL), wherein xiIndicate the color feature vector and texture feature vector structure of i-th of image block At feature vector, yi∈ { 1,2,3 } indicates the type of i-th of image block；The image block of non-marking types is denoted as (xL+1, yL+1),(x L+2,yL+2),…,(xK,yK), the type of these image blocks is unknown；Regard each image block as a vertex, constructs Complete graph connects the weight definition on the side of vertex i and j in figure are as follows:
Wherein, α is super ginseng, preferably 1.5.According to wijSimilar matrix W is constructed, and retains maximum 8 values of every row, Residual value is all set to 0.
Method using label propagation algorithm to image block type is: constructing probability according to similar matrix W first and turns Move matrix P, formula are as follows:
According to there is labeled data (x1,y1),(x2,y2),…,(xL,yL) building two-dimensional matrix M1, size is L × 3, should I-th row of matrix indicates the type vector of i-th of data sample, it may be assumed that if the affiliated type of i-th of data sample is j, jth Column element value is 1, remaining train value is 0；According to no labeled data (xL+1,yL+1),(xL+2,yL+2),…,(xK,yK) building two dimension Matrix M2, size are (K-L) × 3, and each position element initial value of the matrix is disposed as 1/3；By two squares of M1 and M2 Battle array, which merges, obtains matrix F={ M1；M2 }, size is K × 3；Finally carry out label propagation, the method is as follows:
A) it executes label to propagate: F=PF；
B) it resets the label for having labeled data sample in F: L row preceding in F being reset into M1；
C) step a) is repeated and b) until convergence.
The submatrix of the composition of K-L row is the type for indicating not marking image block after in F.
Above-mentioned steps 7) in merge image block detailed process be: calculate all adjacent image blocks first to D；For Image block { Bi,Bj∈ D, B is obtained according to step 8)iAnd BjType, do not handled if type difference, otherwise according to step 3) B is calculatediAnd BjBetween boundary characteristic, if boundary characteristic be less than preset threshold if merge two image blocks, otherwise It is not processed.
The method of the present invention is based on sandstone cross-polarized light microsection image and single polarized light microscopy Slice Image, at image Image pre-segmentation is image block, color characteristic and texture is extracted in cross-polarized light image by reason technology and machine learning method Feature extracts boundary characteristic in single polarized light image；Training Support Vector Machines are right for quartz, feldspar and the cutting grain of classifying Pre-segmentation image block carries out type prediction, and is further processed the fuzzy image block of type using label propagation algorithm；Last root Region merging technique is carried out according to image block type, realizes the automatic segmentation of Sandstone Slice micro-image.The method of the present invention segmentation effect Good, scalability is strong, calculates simplicity, high-efficient, can help geology practitioner that Sandstone Slice micro-image point is effectively reduced It cuts the time and improves the accuracy rate of result；There is the valence of applying in terms of Sandstone Slice identification, oil gas reservoir Layer Detection and assessment Value.
Detailed description of the invention
Fig. 1 is the overall framework of Sandstone Slice micro-image automatic division method；
Fig. 2 is sandstone cross-polarized light microsection image and single polarized light microscopy Slice Image；
Fig. 3 is the process flow diagram that pre-segmentation is carried out to Sandstone Slice micro-image.
Specific embodiment
The main object of the present invention is the automatic segmentation realized to Sandstone Slice micro-image, with image processing techniques and Machine learning method, using super-pixel algorithm by image pre-segmentation as image block, it is special from cross-polarized light image zooming-out color and texture Sign extracts boundary characteristic from single polarized light image；With three kinds of quartz, feldspar and landwaste particle image Training Support Vector Machines, to figure As block type is predicted, and it is further processed with label propagation algorithm；By the neighbouring relations of image block, type and Boundary characteristic carries out image merged block, realizes that Sandstone Slice micro-image is divided automatically.
Fig. 1 show the technological frame of Sandstone Slice micro-image automatic classification method.The input of method is Sandstone Slice Micro-image (including cross-polarized light image and single polarized light image)；The output of method is the segmentation knot of Sandstone Slice micro-image Fruit.For the correct application of ensuring method, the mineral grain image of previously prepared band mark, including quartz particles image, length are needed Stone particle image and cutting grain image, as training sample set.Technological frame is divided into 7 steps: using super-pixel segmentation skill The cross-polarized light microsection image pre-segmentation of input is image block by art；Extract image block color characteristic and textural characteristics； Based on single polarized light microscopy Slice Image, image boundary feature is extracted；Based on mineral grain sample data set Training Support Vector Machines Classifier；The probability that each image block belongs to quartz, feldspar, landwaste is calculated, image block type is determined by preset condition；Make It is the image block marking types of non-marking types with label propagation algorithm；Type is identical, boundary characteristic is less than the adjacent of threshold value Image block merges, and forms complete mineral grain.
Fig. 2 show the made cross-polarized light micro-image of same Sandstone Slice and single resonance offset effect. In cross-polarized light micro-image, the interference colours of different colours are presented in mineral grain, and the adjacent mineral grain in part is compared due to color It is close, thus between obscurity boundary, it is difficult to divide；In addition, there are more complex crystal microstructures in feldspar particle, such as: double , there are a large amount of trifling mineral chips in brilliant, cleavage and rift etc., these micro-structures are often erroneously interpreted as in cutting grain Grain boundary and generate mistake segmentation.In single resonance offset effect, mineral grain be mostly it is colorless or clear, can be therefrom Part mineral grain boundary is extracted, and since the pseudo- boundary that crystal microstructure generates shows inside feldspar particle and cutting grain As unobvious.Therefore, it can be improved the accuracy rate of Sandstone Slice micro-image in conjunction with cross-polarized light image and single polarized light image.
Therefore, the present invention proposes to be based on super-pixel algorithm pre-segmentation sandstone image, in conjunction with cross-polarized light image and single polarisation Image zooming-out image block characteristics, and training classifier identifies image block type, is calculated using label propagation identification image block is failed Method is further processed, to reach better segmentation effect.The step of present invention uses is as follows:
1) sandstone cross-polarized light microsection image is read, uses super-pixel segmentation technology by its pre-segmentation for image block；
2) it is based on cross-polarized light microsection image, extracts image block color characteristic and textural characteristics；
3) sandstone list polarized light microscopy Slice Image is read, image boundary is extracted using bound test technology；In conjunction with pre-segmentation Image block extracts image block boundaries feature；
4) it is based on sandstone particle specimens data set Training Support Vector Machines classifier；
5) trained listening group is used, the probability that each image block belongs to quartz, feldspar, landwaste is calculated；By default Threshold value determines image block classification, higher than the image block marking types of threshold value, remaining image block not marking types；
6) similar matrix is constructed according to the color characteristic of image block and textural characteristics, the use of label propagation algorithm is step 5) The image block marking types of non-marking types；
7) the adjacent image block that type is identical, boundary characteristic is less than threshold value merges, and forms complete mineral grain.
Above-mentioned steps 1) in by image pre-segmentation be image block detailed process be: first reading sandstone cross-polarized light it is micro- Image, along image x-axis and y-axis both direction withFor interval (wherein number of pixels in N image), image is divided At K square area, each region is considered as an initial clustering；Then according to pixel color feature and coordinate value, iteratively It calculates it and belongs to cluster；After iteration ends, the cluster that each pixel is constituted is an image block.
It is as follows to calculate the method clustered belonging to each pixel: according to K initial clustering, calculating the center of each cluster {ck}k∈[1,K], and with ckCentered on search for the set PS of all pixels around it within the scope of 2H × 2H；To pixel each in PS Point i calculates its distance center ckDistance, update the distance if the distance is less than the distance of last iteration, and being arranged should Cluster belonging to pixel is k；After primary to all cluster centre iteration, each cluster centre and new and old cluster are recalculated The Euclidean distance at center stops algorithm if distance is less than preset threshold, otherwise enters next iteration.
Above-mentioned calculating two pixels i and ckBetween distance method it is as follows:
Wherein rc、gcAnd bcFor ckRGB color characteristic value, xcAnd ycFor ckX-axis and y-axis coordinate；ri、giAnd biFor picture The RGB color characteristic value of vegetarian refreshments i, xiAnd yiFor the x-axis and y-axis coordinate of pixel i, m is preferably 8.
Above-mentioned steps 2) in calculate image block color characteristic detailed process be: image block is converted into grayscale image first Picture calculates the color characteristic hist based on histogram, and the feature based on statistics, comprising: mean value Mean, variance Five Variance, intermediate value Median, mode Mode and mean absolute deviation MAD statistical indicators.
Color characteristic hist calculation method based on histogram is as follows: 0~255 gray scale is divided into 16 rank (each grades Bao Han 16 gray scales), calculate the number of pixels { cnt that each grey level is included in image block1,cnt2,...,cnt16, And the vector is normalized to obtain hist={ hist1,hist2,...,hist16}；
Feature calculation formula based on statistics is as follows, the number of pixels for wherein including in N representative image block, IpFor pixel p Gray value:
Median=mid (I1,I2,…,IN), wherein mid () is median function
By above-mentioned calculating, the color characteristic of each image block is the feature vector of 20 dimensions.
Above-mentioned steps 2) in calculate image block textural characteristics detailed process be: image block is converted into grayscale image first Picture, and calculate gray level co-occurrence matrixes；Textural characteristics are calculated according to gray level co-occurrence matrixes.
The method for calculating image grayscale co-occurrence matrix is as follows: arbitrarily taking a point p1=(x, y) in the picture and deviates its One point p2=(x+a, y+b) calculates point to the gray value (f1, f2) of (p1, p2)；It allows point p1 to move in entire image, obtains Different (f1, f2) values；Count the number of each (f1, f2) value appearance；Matrix P is constructed, P (i, j) is defined as gray value The number that (i, j) occurs；Matrix P is normalized, i.e., total time occurred element each in matrix divided by all gray values Number.Define (a, b) respectively (0,1), (1,0), (1,1) and (- 1,1), therefore obtain 4 gray level co-occurrence matrixes.
According to gray level co-occurrence matrixes calculate textural characteristics method it is as follows: calculate separately ENERGY E nergy, entropy Entropy, Maximum probability MaxProbability, contrast C ontrast, reciprocal difference square IDM index, formula are as follows:
The textural characteristics that each gray matrix obtains include 5 elements, therefore the textural characteristics of each image block are 20 dimensions Feature vector.
Above-mentioned steps 3) in using bound test technology extract image boundary detailed process be: first reading sandstone list it is inclined Light microsection image converts it to gray level image, and pre-processes to image；Then using Canny operator detection figure Boundary as in；Finally, calculating the boundary characteristic of image block according to image boundary.
It is as follows that pretreated method is carried out to image: image being smoothed using Gaussian filter first；Then Gamma correction is carried out to image；Finally carry out histogram equalization.
The method for carrying out gamma correction to image is: for pixel p, enabling its gray value is Ip, first to IpNormalization, Formula are as follows: x=(Ip+0.5)/256；Then x is pre-compensated for, formula are as follows: y=x1/γ, select γ=2.2；Finally carry out Renormalization, formula are as follows: z=y*256-0.5, z are the gray value of pixel p after gamma correction.
The method for carrying out histogram equalization to image is: enabling in image has N number of pixel, and image grayscale is divided into l Gray level, l are preferably 16, and calculating gray level is rkPixel number nk, then the probability that k-th of gray level occurs is P (rk) =nk/ N, k=0,1 ..., l-1, histogram equalization formula are as follows:
Wherein skBe gray level be k pixel through transformation after gray scale.
Above-mentioned steps 4) in the detailed process of Training Support Vector Machines classifier be: firstly, collecting quartz, feldspar, landwaste The image sample data collection of particle；The quantity for counting every class sandstone particle image implements image to the grain type of negligible amounts The method of enhancing is generated additional image pattern by the methods of translation, overturning, rotation, scaling, makes three types particle figure Picture balances between reaching class；For each sandstone particle specimens image, color characteristic and texture are calculated according to the method for step 2) Feature, and constitutive characteristic vector；It is marked according to particle image classification, constructs classifier training collection；Finally, based on image training Collection, Training Support Vector Machines classifier.
Above-mentioned steps 5) in be to the detailed process that image block type is labeled: using the trained support of step 4) The color feature vector and texture feature vector that step 2) is calculated in vector machine classifier predict, obtain probability to Measure p={ p1,p2,p3, it is the probability that the image block belongs to quartz, feldspar and landwaste respectively；Enable pmax=max (p1,p2, p3), As pmax >=0.7, the type for marking image block is the type of sandstone particle corresponding to pmax.
Above-mentioned steps 6) in treatment process that the type for not marking image block is predicted be: first by step 1) To each image block regard a vertex as, construct similar matrix；Then using label propagation algorithm to the figure of non-marking types As block carries out type mark.
The method of building similar matrix is: having the image block of marking types to constitute labeled data step 5), is denoted as (x1, y1),(x2,y2),…,(xL,yL), wherein xiIndicate the color feature vector and texture feature vector structure of i-th of image block At feature vector, yi∈ { 1,2,3 } indicates the type of i-th of image block；The image block of non-marking types is denoted as (xL+1, yL+1),(x L+2,yL+2),…,(xK,yK), the type of these image blocks is unknown；Regard each image block as a vertex, constructs Complete graph connects the weight definition on the side of vertex i and j in figure are as follows:
Wherein, α is super ginseng, preferably 1.5.According to wijSimilar matrix W is constructed, and retains maximum 8 values of every row, Residual value is all set to 0.
Method using label propagation algorithm to image block type is: constructing probability according to similar matrix W first and turns Move matrix P, formula are as follows:
According to there is labeled data (x1,y1),(x2,y2),…,(xL,yL) building two-dimensional matrix M1, size is L × 3, should I-th row of matrix indicates the type vector of i-th of data sample, it may be assumed that if the affiliated type of i-th of data sample is j, jth Column element value is 1, remaining train value is 0；According to no labeled data (xL+1,yL+1),(xL+2,yL+2),…,(xK,yK) building two dimension Matrix M2, size are (K-L) × 3, and each position element initial value of the matrix is disposed as 1/3；By two squares of M1 and M2 Battle array, which merges, obtains matrix F={ M1；M2 }, size is K × 3；Finally carry out label propagation, the method is as follows:
A) it executes label to propagate: F=PF；
B) it resets the label for having labeled data sample in F: L row preceding in F being reset into M1；
C) step a) is repeated and b) until convergence.
The submatrix of the composition of K-L row is the type for indicating not marking image block after in F.
Above-mentioned steps 7) in merge image block detailed process be: calculate all adjacent image blocks first to D；For Image block { Bi,Bj∈ D, B is obtained according to step 8)iAnd BjType, do not handled if type difference, otherwise according to step 3) B is calculatediAnd BjBetween boundary characteristic, if boundary characteristic be less than preset threshold if merge two image blocks, otherwise It is not processed.
The method of the present invention makes full use of the characteristics of sandstone cross-polarized light image and single polarized light image, extracts its color spy respectively Sign, textural characteristics and boundary characteristic, application image processing technique, machine learning method realize oneself of Sandstone Slice micro-image Dynamic segmentation；Aiming at the problem that, obscurity boundary close adjacent particle feature in Sandstone Slice Image, the present invention combines single polarisation figure As assisted border detection；For the problem that crystal microstructure is more complex in feldspar and cutting grain, segmentation is influenced, the present invention is logical It crosses extraction cross-polarized light color of image and textural characteristics classifies to image block, and combination tag propagation algorithm is further true Determine the affiliated mineral grain type of image block, the generation that crystal microstructure is accidentally divided into boundary is avoided, to reach ideal segmentation Effect.By the Sandstone Slice micro-image from the acquisition of the ground such as Tibet and production as experimental data, the experimental results showed that this hair Bright method can reach wanting substantially for thin section identification to the segmentation of various types of Sandstone Slices accuracy rate with higher It asks.In addition, the method for the present invention can be applied to other kinds of sedimentary rock, and as: shale, limestone thin slice micro-image Segmentation has preferable scalability.
There are many concrete application approach of the method for the present invention, the above is only a preferred embodiment of the present invention.It should refer to Out, for those skilled in the art, without departing from the principle of the present invention, if can also make Dry to improve, these improvement also should be regarded as protection scope of the present invention.

Claims (9)

1. a kind of automatic division method of sandstone microsection image, it is characterised in that the following steps are included:
1) sandstone cross-polarized light microsection image is read, uses super-pixel segmentation technology by its pre-segmentation for image block；
2) it is based on cross-polarized light microsection image, extracts image block color characteristic and textural characteristics；
3) sandstone list polarized light microscopy Slice Image is read, image boundary is extracted using bound test technology；In conjunction with pre-segmentation image Block extracts image block boundaries feature；
4) it is based on sandstone particle specimens data set Training Support Vector Machines classifier；
5) trained listening group is used, the probability that each image block belongs to quartz, feldspar, landwaste is calculated；Pass through preset threshold Determine image block classification, higher than the image block marking types of threshold value, remaining image block not marking types；
6) similar matrix is constructed according to the color characteristic of image block and textural characteristics, the use of label propagation algorithm is that step 5) is not marked Infuse the image block marking types of type；
7) the adjacent image block that type is identical, boundary characteristic is less than threshold value merges, and forms complete mineral grain.
2. the automatic division method of sandstone microsection according to claim 1, which is characterized in that above-mentioned steps 1) place Reason process is: reading sandstone cross-polarized light micro-image first divides the image into K square area, each region is considered as one A initial clustering；Then it according to pixel color feature and coordinate value, iteratively calculates it and belongs to cluster；After iteration ends, each picture The cluster that vegetarian refreshments is constituted is an image block.
The method for dividing the image into K square area is as follows: the number of pixels for enabling image include is N, along image x-axis and y-axis Both direction withFor interval, image uniform is divided into K square area (initial clustering).
It is as follows to calculate the method clustered belonging to each pixel: according to K initial clustering, calculating the center of each cluster {ck}k∈[1,K], and with ckCentered on search for the set PS of all pixels around it within the scope of 2H × 2H；To pixel each in PS Point i calculates its distance center ckDistance, update the distance if the distance is less than the distance of last iteration, and being arranged should Cluster belonging to pixel is k；After primary to all cluster centre iteration, recalculate in each cluster centre and new and old cluster The Euclidean distance of the heart stops algorithm if distance is less than preset threshold, otherwise enters next iteration.
Above-mentioned calculating two pixels i and ckBetween distance method it is as follows:
Wherein rc、gcAnd bcFor ckRGB color characteristic value, xcAnd ycFor ckX-axis and y-axis coordinate；ri、giAnd biFor pixel i RGB color characteristic value, xiAnd yiFor the x-axis and y-axis coordinate of pixel i, m is preferably 8.
3. the automatic division method of sandstone microsection according to claim 2, which is characterized in that above-mentioned steps 2) it falls into a trap The treatment process for calculating image block color characteristic is: image block being converted to gray level image first, calculates the color based on histogram Feature hist, and the feature based on statistics, comprising: mean value Mean, variance Variance, intermediate value Median, mode Mode and Five statistical indicators of mean absolute deviation MAD.
Color characteristic hist calculation method based on histogram is as follows: 0~255 gray scale is divided into 16 ranks (each rank packet Containing 16 gray scales), calculate the number of pixels { cnt that each grey level is included in image block1,cnt2,...,cnt16, and will The vector normalizes to obtain hist={ hist1,hist2,...,hist16}；
Feature calculation formula based on statistics is as follows, the number of pixels for wherein including in N representative image block, IpFor the gray scale of pixel p Value:
Median=mid (I1,I2,…,IN), wherein mid () is median function
By above-mentioned calculating, the color characteristic of each image block is the feature vector of 20 dimensions.
4. the automatic division method of sandstone microsection according to claim 3, which is characterized in that above-mentioned steps 2) it falls into a trap The treatment process for calculating image block textural characteristics is: image block being converted to gray level image first, and calculates gray level co-occurrence matrixes；Root Textural characteristics are calculated according to gray level co-occurrence matrixes.
Calculate image grayscale co-occurrence matrix method it is as follows: arbitrarily take in the picture a point p1=(x, y) and deviate it a bit P2=(x+a, y+b) calculates point to the gray value (f1, f2) of (p1, p2)；It allows point p1 to move in entire image, obtains difference (f1, f2) value；Count the number of each (f1, f2) value appearance；Matrix P is constructed, P (i, j) is defined as gray value (i, j) Existing number；Matrix P is normalized, i.e., the total degree occurred element each in matrix divided by all gray values.Definition (a, b) is respectively (0,1), (1,0), (1,1) and (- 1,1), therefore obtain 4 gray level co-occurrence matrixes.
It is as follows according to the method that gray level co-occurrence matrixes calculate textural characteristics: to calculate separately ENERGY E nergy, entropy Entropy, maximum Probability MaxProbability, contrast C ontrast, reciprocal difference square IDM index, formula are as follows:
The textural characteristics that each gray matrix obtains include 5 elements, therefore the texture feature vector of each image block includes 20 A element.
5. the automatic division method of sandstone microsection according to claim 4, which is characterized in that above-mentioned steps 3) place Reason process is: reading sandstone list polarized light microscopy Slice Image first converts it to gray level image, and located in advance to image Reason；Then using the boundary in Canny operator detection image；Finally, calculating the boundary characteristic of image block according to image boundary.
It is as follows that pretreated method is carried out to image: image being smoothed using Gaussian filter first；Then to figure As carrying out gamma correction；Finally carry out histogram equalization.
The method for carrying out gamma correction to image is: for pixel p, enabling its gray value is Ip, first to IpNormalization, formula Are as follows: x=(Ip+0.5)/256；Then x is pre-compensated for, formula are as follows: y=x1/γ, select γ=2.2；Finally carry out anti-normalizing Change, formula are as follows: z=y*256-0.5, z are the gray value of pixel p after gamma correction.
The method for carrying out histogram equalization to image is: enabling in image has N number of pixel, and image grayscale is divided into l gray scale Grade, l is preferably 16, and calculating gray level is rkPixel number nk, then the probability that k-th of gray level occurs is P (rk)=nk/ N, k=0,1 ..., l-1, histogram equalization formula are as follows:
Wherein skBe gray level be k pixel through transformation after gray scale.
6. the automatic division method of sandstone microsection according to claim 5, which is characterized in that above-mentioned steps 4) place Reason process is: firstly, collecting the image sample data collection of quartz, feldspar, cutting grain；Count the number of every class sandstone particle image Amount is implemented image enchancing method to the grain type of negligible amounts, is generated by the methods of translation, overturning, rotation, scaling additional Image pattern balances between so that three types particle image is reached class；For each sandstone particle specimens image, according to step 2) Method calculate color characteristic and textural characteristics, and constitutive characteristic vector；It is marked according to particle image classification, building classifier instruction Practice collection；Finally, being based on training set of images, Training Support Vector Machines classifier.
7. the automatic division method of sandstone microsection according to claim 6, which is characterized in that above-mentioned steps 5) place Reason process is: the color feature vector and line that step 2) is calculated using step 4) trained support vector machine classifier Reason feature vector is predicted, probability vector p={ p is obtained1,p2,p3, it is that the image block belongs to quartz, feldspar and landwaste respectively Probability；Enable pmax=max (p1,p2,p3), as pmax >=0.7, the type for marking image block is sandstone corresponding to pmax The type of particle.
8. the automatic division method of sandstone microsection according to claim 7, which is characterized in that above-mentioned steps 6) place Reason process is: each image block for first obtaining step 1) regards a vertex as, constructs similar matrix；Then it is passed using label It broadcasts algorithm and type mark is carried out to the image block of non-marking types.
The method of building similar matrix is: the image block that step 5) marked type being constituted labeled data, is denoted as (x1,y1), (x2,y2),…,(xL,yL), wherein xiIndicate the color feature vector of i-th of image block and the feature that texture feature vector is constituted Vector, yi∈ { 1,2,3 } indicates the type of i-th of image block；The image block of non-marking types is denoted as (xL+1,yL+1),(x L+2, yL+2),…,(xK,yK), the type of these image blocks is unknown；Regard each image block as a vertex, construct complete graph, in figure Connect the weight definition on the side of vertex i and j are as follows:
Wherein, α is super ginseng, preferably 1.5.According to wijSimilar matrix W is constructed, and retains maximum 8 values of every row, residual value is complete Portion is set as 0.
Method using label propagation algorithm to image block type is: constructing probability according to similar matrix W first and shifts square Battle array P, formula are as follows:
According to there is labeled data (x1,y1),(x2,y2),…,(xL,yL) building two-dimensional matrix M1, size is L × 3, the matrix The i-th row indicate i-th of data sample type vector, it may be assumed that if the affiliated type of i-th of data sample be j, jth column member Element value is 1, remaining train value is 0；According to no labeled data (xL+1,yL+1),(xL+2,yL+2),…,(xK,yK) building two-dimensional matrix M2, size are (K-L) × 3, and each position element initial value of the matrix is disposed as 1/3；Two matrixes of M1 and M2 are closed And obtain matrix F={ M1；M2 }, size is K × 3；Finally carry out label propagation, the method is as follows:
A) it executes label to propagate: F=PF；
B) it resets the label for having labeled data sample in F: L row preceding in F being reset into M1；
C) step a) is repeated and b) until convergence.
The submatrix of the composition of K-L row is the type for indicating not marking image block after in F.
9. the automatic division method of sandstone microsection according to claim 8, which is characterized in that above-mentioned steps 7) place Reason process is: calculating all adjacent image blocks first to D；For image block { Bi,Bj∈ D, B is obtained according to step 8)iAnd Bj Type, do not handled if type difference, B be otherwise calculated according to step 3)iAnd BjBetween boundary characteristic, if side Boundary's feature is less than preset threshold and then merges two image blocks, is otherwise not processed.
CN201811085809.3A 2018-09-18 2018-09-18 A kind of automatic division method of Sandstone Slice micro-image Pending CN109523566A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811085809.3A CN109523566A (en) 2018-09-18 2018-09-18 A kind of automatic division method of Sandstone Slice micro-image

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811085809.3A CN109523566A (en) 2018-09-18 2018-09-18 A kind of automatic division method of Sandstone Slice micro-image

Publications (1)

Publication Number Publication Date
CN109523566A true CN109523566A (en) 2019-03-26

Family

ID=65771344

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811085809.3A Pending CN109523566A (en) 2018-09-18 2018-09-18 A kind of automatic division method of Sandstone Slice micro-image

Country Status (1)

Country Link
CN (1) CN109523566A (en)

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111008949A (en) * 2019-08-16 2020-04-14 苏州喆安医疗科技有限公司 Soft and hard tissue detection method for tooth image
CN111007064A (en) * 2019-12-13 2020-04-14 常州大学 Intelligent logging lithology identification method based on image identification
CN111353976A (en) * 2020-02-25 2020-06-30 南京大学 Sand grain target detection method based on convolutional neural network
CN111413284A (en) * 2020-05-18 2020-07-14 攀钢集团攀枝花钢铁研究院有限公司 Quantitative detection method for main phase in vanadium-titanium sinter
CN111709423A (en) * 2020-06-26 2020-09-25 东北石油大学 Rock debris lithology recognition method under PDC drill bit condition matched with lithology feature library
CN112784894A (en) * 2021-01-18 2021-05-11 西南石油大学 Automatic labeling method for rock slice microscopic image
CN112991370A (en) * 2021-04-08 2021-06-18 重庆大学 Rock core CT crack identification and segmentation method
CN113139936A (en) * 2021-03-31 2021-07-20 武汉钢铁有限公司 Image segmentation processing method and device
CN113378825A (en) * 2021-07-09 2021-09-10 中海石油(中国)有限公司 Sandstone slice image identification method and system based on artificial intelligence
CN113435458A (en) * 2021-02-08 2021-09-24 中国石油化工股份有限公司 Rock slice image segmentation method, device and medium based on machine learning
CN113686904A (en) * 2021-07-28 2021-11-23 河北工业大学 Method for describing micro-macroscopic fracture and micro-crack formation of rock mass under external load
CN114897917A (en) * 2022-07-13 2022-08-12 西南石油大学 Multi-level rock casting body slice image segmentation method
CN116205908A (en) * 2023-04-27 2023-06-02 阳谷新太平洋电缆有限公司 Cable coaxiality visual detection method based on convolutional neural network

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106485223A (en) * 2016-10-12 2017-03-08 南京大学 The automatic identifying method of rock particles in a kind of sandstone microsection
CN106952393A (en) * 2017-03-28 2017-07-14 深圳怡化电脑股份有限公司 Paper Currency Identification and device, electronic equipment and storage medium
CN107146233A (en) * 2017-04-24 2017-09-08 四川大学 Granulometry Segmentation based on petrographic thin section polarisation sequence chart

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106485223A (en) * 2016-10-12 2017-03-08 南京大学 The automatic identifying method of rock particles in a kind of sandstone microsection
CN106952393A (en) * 2017-03-28 2017-07-14 深圳怡化电脑股份有限公司 Paper Currency Identification and device, electronic equipment and storage medium
CN107146233A (en) * 2017-04-24 2017-09-08 四川大学 Granulometry Segmentation based on petrographic thin section polarisation sequence chart

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
FENG JIANG 等: ""Grain segmentation of multi-angle petrographic thin section microscopic images"", 《2017 IEEE INTERNATIONAL CONFERENCE ON IMAGE PROCESSING (ICIP)》 *
SNEHAMOY CHATTERJEE 等: ""Genetic algorithms for feature selection of image analysis-based quality monitoring model: An application to an iron mine"", 《ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE》 *
卢影: ""基于标签传播的半监督多标签学习"", 《中国优秀硕士学位论文全文数据库信息科技辑》 *
郝慧珍 等: ""砂岩显微图像分析方法及其工具实现"", 《计算机科学》 *

Cited By (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111008949B (en) * 2019-08-16 2021-09-14 苏州喆安医疗科技有限公司 Soft and hard tissue detection method for tooth image
CN111008949A (en) * 2019-08-16 2020-04-14 苏州喆安医疗科技有限公司 Soft and hard tissue detection method for tooth image
CN111007064A (en) * 2019-12-13 2020-04-14 常州大学 Intelligent logging lithology identification method based on image identification
CN111353976A (en) * 2020-02-25 2020-06-30 南京大学 Sand grain target detection method based on convolutional neural network
CN111413284A (en) * 2020-05-18 2020-07-14 攀钢集团攀枝花钢铁研究院有限公司 Quantitative detection method for main phase in vanadium-titanium sinter
CN111709423A (en) * 2020-06-26 2020-09-25 东北石油大学 Rock debris lithology recognition method under PDC drill bit condition matched with lithology feature library
CN111709423B (en) * 2020-06-26 2022-03-15 东北石油大学 Rock debris lithology recognition method under PDC drill bit condition matched with lithology feature library
CN112784894A (en) * 2021-01-18 2021-05-11 西南石油大学 Automatic labeling method for rock slice microscopic image
CN113435458A (en) * 2021-02-08 2021-09-24 中国石油化工股份有限公司 Rock slice image segmentation method, device and medium based on machine learning
WO2022166232A1 (en) * 2021-02-08 2022-08-11 中国石油化工股份有限公司 Rock identification method, system and apparatus, terminal, and readable storage medium
CN113139936A (en) * 2021-03-31 2021-07-20 武汉钢铁有限公司 Image segmentation processing method and device
CN113139936B (en) * 2021-03-31 2022-07-08 武汉钢铁有限公司 Image segmentation processing method and device
CN112991370A (en) * 2021-04-08 2021-06-18 重庆大学 Rock core CT crack identification and segmentation method
CN112991370B (en) * 2021-04-08 2022-11-25 重庆大学 Rock core CT crack identification and segmentation method
CN113378825A (en) * 2021-07-09 2021-09-10 中海石油(中国)有限公司 Sandstone slice image identification method and system based on artificial intelligence
CN113378825B (en) * 2021-07-09 2024-04-05 中海石油(中国)有限公司 Sandstone sheet image identification method and system based on artificial intelligence
CN113686904A (en) * 2021-07-28 2021-11-23 河北工业大学 Method for describing micro-macroscopic fracture and micro-crack formation of rock mass under external load
CN113686904B (en) * 2021-07-28 2023-08-29 河北工业大学 Method for describing microscopic fracture and microcrack formation of rock mass under external load
CN114897917A (en) * 2022-07-13 2022-08-12 西南石油大学 Multi-level rock casting body slice image segmentation method
CN116205908A (en) * 2023-04-27 2023-06-02 阳谷新太平洋电缆有限公司 Cable coaxiality visual detection method based on convolutional neural network

Similar Documents

Publication Publication Date Title
CN109523566A (en) A kind of automatic division method of Sandstone Slice micro-image
CN106127725B (en) A kind of millimetre-wave radar cloud atlas dividing method based on multiresolution CNN
CN103971115B (en) Automatic extraction method for newly-increased construction land image spots based on NDVI and PanTex index
CN106485223B (en) The automatic identifying method of rock particles in a kind of sandstone microsection
CN109919206A (en) A kind of remote sensing image ground mulching classification method based on complete empty convolutional neural networks
CN108416353A (en) Crop field spike of rice fast partition method based on the full convolutional neural networks of depth
CN106529508A (en) Local and non-local multi-feature semantics-based hyperspectral image classification method
CN107392130A (en) Classification of Multispectral Images method based on threshold adaptive and convolutional neural networks
CN104134068B (en) Monitoring vehicle feature representation and classification method based on sparse coding
CN107832797B (en) Multispectral image classification method based on depth fusion residual error network
CN105513066B (en) It is a kind of that the generic object detection method merged with super-pixel is chosen based on seed point
CN104217440B (en) A kind of method extracting built-up areas from remote sensing images
CN107358176A (en) Sorting technique based on high score remote sensing image area information and convolutional neural networks
CN105354600B (en) A kind of automatic classification method of sandstone microsection
CN104239890A (en) Method for automatically extracting coastal land and earth cover information through GF-1 satellite
CN110263610A (en) A kind of degeneration file and picture binary coding method and system based on deep learning
Gamba et al. Urban climate zone detection and discrimination using object-based analysis of VHR scenes
Chen et al. Land cover mapping in urban environments using hyperspectral APEX data: A study case in Baden, Switzerland
CN106557758B (en) A kind of multiple target automatic identification method of grains of sand micro-image
CN105243387A (en) Open-pit mine typical ground object classification method based on UAV image
CN108154158A (en) A kind of building image partition method applied towards augmented reality
CN106529600B (en) A kind of high-resolution optical image building corner recognition method based on SVM
CN102231190B (en) Automatic extraction method for alluvial-proluvial fan information
Honawad et al. Soil classification and suitable crop prediction
Kupková et al. Towards reliable monitoring of grass species in nature conservation: Evaluation of the potential of UAV and PlanetScope multi-temporal data in the Central European tundra

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20190326

RJ01 Rejection of invention patent application after publication