CN102855492A - Classification method based on mineral flotation foam image - Google Patents

Abstract

The invention discloses a classification method based on a mineral flotation foam image. A real-time acquired foam image is classified in different known working conditions. The classification method comprises the following steps of: introducing a vocabulary in text classification into a flotation foam image, blocking the foam image acquired by an industrial camera and extracting the characteristic parameters, clustering the color and texture characteristic parameters of the extracted foam image by employing a K-mean clustering method, obtaining a plurality of clustering centers, and constructing a foam state vocabulary; describing the real-time foam image through a word bag method by utilizing the obtained foam state vocabulary, and forming a vector representation of the foam image; and finally, classifying the foam image through the similarity between measuring vectors by employing a vector space model. Because different types correspond to different working conditions, the flotation working condition recognition can be performed according to the classification result of the foam image; and therefore, the operation guide is given, the production is optimized, and the production efficiency is improved.

Description

Sorting technique based on the mineral floating froth images

[technical field]

The invention belongs to the mineral floating field.The sorting technique of mineral floating froth images particularly.

[background technology]

The mineral floating production run is generally controlled by the state that experienced workman observes flotation froth, and the uncertainty of this operation is difficult to make floatation process to operate in optimum state.Utilizing digital image processing techniques, floatation foam image is classified and explained, obtain the work information of mineral floating, and then be optimized control, is a kind of effective ways of increasing economic efficiency.Present stage mainly is that low-level image feature parameters such as textural characteristics, color characteristic, foam size distribution by extracting floatation foam image are described froth images to the research of the classification of floatation foam image and identification, then utilize the methods such as neural network or support vector machine to carry out Classification and Identification, produce to instruct thereby obtain operating mode.Yet, image is done as a whole the processing, the froth images of only being described with the image low-level image feature is not easy to people to be understood, and namely there is the semantic gap problem in the classification results based on the image low-level image feature.In addition, because the local message of froth images does not obtain utilizing, and there are the noise such as illumination, dust in industry spot, two diverse images, the overall low-level image feature that extracts is described may be very approaching, this has just caused based on the accuracy of the Classification and Identification of the neural network of froth images low-level image feature or SVM low, thereby makes based on the frequent misoperation of the production operation of operating mode, produces to be difficult to stablize optimized operation.

Vector space model (VSM) is proposed in 20 century 70s by people such as Salton, is a kind of algebraic model that represents text.VSM is the vector operation that the processing of content of text is reduced in the vector space, and expresses semantic similarity with the similarity on the space, and is visual and understandable.Corresponding to froth images, if can remove to understand image from semantic level, on semantic level, froth images is carried out Classification and Identification again, then classification results is more near people's understanding.Because there is the impact of the illumination, dust of industry spot etc. in froth images, and all kinds of froth images itself has, and brightness is low, the similarity high, removes to distinguish all kinds of images from details, part and just seems more important.

[summary of the invention]

The object of the invention is to have semantic gap and the inaccurate problem of classification for the image classification method of describing based on the flotation froth low-level image feature, a kind of sorting technique based on the mineral floating froth images is provided.

Floatation foam image sorting technique provided by the present invention mainly comprises three phases: (1) generates based on the froth images foam state vocabulary of textural characteristics and color characteristic; (2) the word bag of froth images is described; (3) utilize vector space model to carry out the froth images classification.Specifically describe as follows:

(1) the froth images foam state vocabulary based on textural characteristics and color characteristic generates

On-the-spot for flotation, the froth images of different operating modes has very large difference, and the depth of fineness degree, groove that is mainly manifested in texture is not equal, in addition, flotation froth has obvious color, color can reflect to a great extent the entrained mineral type of foam, mineral what etc. information.Therefore select the textural characteristics of froth images and color characteristic to come froth images is described.Utilize gray level co-occurrence matrixes (GLCM) algorithm to extract the parametric texture of froth images, comprise angle second moment, entropy, contrast, unfavourable balance square and correlativity; Then extract the color characteristic of froth images, i.e. relative red component.

Gray level co-occurrence matrixes is by the joint probability density P(i between the image gray levels, j, d, θ) matrix that consists of, reflected the image spatial coherence of gray scale between any two points from the angle of statistics.Its definition direction is that θ, spacing are the gray level co-occurrence matrixes P(i of d, j, d, θ) be the value of the capable j column element of i of co-occurrence matrix.θ gets 0 ⁰, 45 ⁰, 90 ⁰With 135 ⁰4 directions are established the digital picture that I (x, y) is a width of cloth two dimension, and its size is U * V pixel, and x, y are respectively the pixel coordinate value of pixel.Then for different θ, P(i, j, d, θ) computing method are as follows:

P(i,j,d,0°)=Num{(x ₁,y ₁)(x ₂,y ₂)∈U*V|x ₁-x ₂=0,

|y ₁-y ₂|=d;I(x ₁,y ₁)=i,I(x ₂,y ₂)=j}

P (i, j, d, 45 °)=Num{ (x ₁, y ₁) (x ₂, y ₂) ∈ U*V| (x ₁-x ₂=d, y ₁-y ₂=d) or

(x ₁-x ₂=-d,y ₁-y ₂=-d);I(x ₁,y ₁)=i,I(x ₂,y ₂)=j}

P(i,j,d,90°)=Num{(x ₁,y ₁)(x ₂,y ₂)∈U*V|x ₁-x ₂=d,

y ₁-y ₂=0;I(x ₁,y ₁)=i,I(x ₂,y ₂)=j}

P (i, j, d, 135 °)=Num{ (x ₁, y ₁) (x ₂, y ₂) ∈ U*V| (x ₁-x ₂=d, y ₁-y ₂=-d) or

(x ₁-x ₂=-d,y ₁-y ₂=d);I(x ₁,y ₁)=i,I(x ₂,y ₂)=j}

Wherein Num{X} represents to gather the element number among the X.

Texture, color parameter computing formula and be described below:

1. ASM (angle second moment)

$\mathrm{ASM}=\underset{i,j}{\mathrm{\Σ}}{\left\{P(i,j|d,\mathrm{\θ})\right\}}^2$

The gradation uniformity of Description Image, texture is thick, and ASM is larger, otherwise then less.

2. ENT (entropy)

$\mathrm{ENT}=-\underset{i,j}{\mathrm{\Σ}}\left\{P(i,j|d,\mathrm{\θ})\right\}\log \{P(i,j|d,\mathrm{\θ})\}$

The quantity of information that Description Image has, if the texture of image is many, entropy is just large; Otherwise texture is few, is exactly smoother, and then entropy is just little.

3. CON (contrast)

$\mathrm{CON}=\mathrm{\Σ}{(i-j)}^{2}P(i,j|d,\mathrm{\θ})$

The sharpness of Description Image, texture rill depth degree, rill is darker, and contrast is larger, and image is also more clear.

4. IDM (unfavourable balance square)

$\mathrm{IDM}=\underset{i,j}{\mathrm{\Σ}}P(i,j|d,\mathrm{\θ})/[1+{(i-j)}^2]$

The homogeney of Description Image texture, what of tolerance image texture localized variation.Co-occurrence matrix is along in the diagonal set, and then unfavourable balance square value is larger.

5. COR (correlativity)

$\mathrm{COR}=\underset{i,j}{\mathrm{\Σ}}(i-{\mathrm{\μ}}_x)(j-{\mathrm{\μ}}_y)P(i,j|d,\mathrm{\θ})/{\mathrm{\σ}}_x{\mathrm{\σ}}_y$

${\mathrm{\μ}}_x=\underset{i}{\mathrm{\Σ}}i\underset{j}{\mathrm{\Σ}}p(i,j|d,\mathrm{\θ})$ ， ${\mathrm{\μ}}_y=\underset{j}{\mathrm{\Σ}}j\underset{i}{\mathrm{\Σ}}p(i,j|d,\mathrm{\θ})$

${\mathrm{\σ}}_x=\underset{i}{\mathrm{\Σ}}{(i-{\mathrm{\μ}}_x)}^2\underset{j}{\mathrm{\Σ}}p(i,j|d,\mathrm{\θ})$ ， ${\mathrm{\σ}}_y=\underset{j}{\mathrm{\Σ}}{(j-{\mathrm{\μ}}_y)}^2\underset{i}{\mathrm{\Σ}}p(i,j|d,\mathrm{\θ})$

Describe the similarity degree of element on the direction of row or column in the gray level co-occurrence matrixes, reflect certain color at the development length along certain direction, extension longer, COR is larger.

6. R _Relative(relatively red component)

$R_{\mathrm{relative}}=\frac{R_{\mathrm{red}}}{R_{\mathrm{gray}}}$

In the formula, R _RedAnd R _GrayRepresent respectively red component average and gray average.

Froth images foam state vocabulary generative process is as follows:

Step 1: choose N width of cloth image (all image category of covering that this N width of cloth image will be wide as far as possible) from image library, every width of cloth image all is truncated to a certain identical pixel size L _x* L _y

Step 2: for the N width of cloth image after the intercepting, every width of cloth image evenly is divided into m * m piece;

Step 3: for each piecemeal, obtain its textural characteristics value and relative red color component value, consist of the low-level image feature vector description of one 1 * 6 dimension;

Step 4: all low-level image feature vector descriptions are carried out the K-means cluster, and the D that an obtains cluster centre is the foam state vocabulary.

(2) the word bag of froth images is described

After obtaining the foam state vocabulary of froth images, method that just can the word bag is described each width of cloth froth images, obtains a vector representation of image.

It is as follows that the word bag of froth images is described process:

Step 1: the image pixel intercepting is L _x* L _yPixel size;

Step 2: image evenly is divided into m * m piece;

Step 3: for each piecemeal, obtain its textural characteristics and relative red color component value, consist of one 1 * 6 dimension low-level image feature vector description;

Step 4: by calculating the Euclidean distance of each foam state vocabulary in this low-level image feature vector description and the foam state vocabulary, measure the similarity of all vocabulary in the vector description of each piecemeal and the foam state vocabulary, the most similar to which foam state vocabulary, which vocabulary just it is demarcated is;

Step 5: add up each foam state vocabulary frequency of occurrence, obtain the word bag vector representation of image.

Described m value is 5 ~ 20.

(3) utilize vector space model to carry out the froth images classification

Have based on the training of vector space model and sorting technique multiple, such as bayes method, k nearest neighbor method, support vector machine, neural net method etc.The present invention classifies to froth images with inner product of vectors classification and two kinds of methods of k nearest neighbor classification respectively.

Method 1: inner product of vectors classification

Stable for classifying quality, the present invention does not characterize such image with the center vector of each classification, but classifies by all Image similarity sums in each classification in calculating image to be classified and the training set.The classification thinking is that for the image in each classification in the training set, the method for word bag is described, and obtains such word bag vector set C _i, i=1,2 ..., N _S(N _SClassification number for classified image in the training set).When getting access to new realtime graphic, the method for this image word bag is described, obtain the vector representation of image, gather C by calculating this word bag vector with the word bag vector of all kinds of images _iSimilarity classify.Detailed process is as follows:

Step 1: will train the image collection word bag model of each classification in the picture library to be described the vector set that obtains such image, each classification image is got M width of cloth image (M value 10 ~ 200), and then the word bag vector set of i class image is combined into

$C_i={\left[\begin{array}{cccc}{c}_{i1}& {\mathrm{<figure-callout\; id="2"\; label="c\; i"\; filenames="2012102651258100003DEST\_PATH\_IMAGE001.png"\; state="\{\{state\}\}">c</figure-callout>}}_i& \&CenterDot;\&CenterDot;\&CenterDot;& c_{\mathrm{iM}}\end{array}\right]}^T,i=\mathrm{1,2},...,N_S$

In the formula, c _IjThe word bag that represents j image of i class is described, and is the row vector of 1 * D;

Step 2: for image to be sorted, obtain its word bag and describe Q, Q=[q ₁q ₂Q _D] be the row vector of 1 * D;

Step 3: with Q and C _iThe vectorial c of in the class each _IjCarry out similarity and calculate (measuring its similarity with the inner product of vector here), obtain Q and each classification set C _iThe similarity sum SC of middle institute directed quantity _i

${\mathrm{SC}}_i=\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}\mathrm{Sim}(c_{\mathrm{ij}},Q),i=\mathrm{1,2,3}...,N_S$

In the formula, $\mathrm{Sim}(W,Q)=W\&CenterDot;Q=\underset{i=1}{\overset{D}{\mathrm{\&Sigma;}}}{w}_i\&times;q_i$ , W=[w ₁w ₂W _D] be the row vector of 1 * D;

Step 4: Q is classified as the k class of similarity maximum,

Method 2:K nearest neighbour classification method

The thinking of this sorting technique is, for image to be sorted, first word bag method is described, and calculation training is concentrated K the image the most similar to this new images, judges classification under the new images according to the classification under K the image.Concrete steps are as follows:

Step 1: N width of cloth training image word bag method is described, obtained a vector set Z as training vector collection, Z=[Z ₁Z ₂Z _N] ^T, Z wherein _i=[z _I1z _I2Z _ID] be the word bag vector description of i width of cloth image.

Step 2: to image to be classified, word bag method is described, and obtains a vectorial Q, Q=[q ₁q ₂Q _D] be the row vector of 1 * D.

Step 3: select K image the most similar to image to be classified Q from training image.The computing formula of similarity is:

$\mathrm{Sim}(Z_i,Q)=Z_i\&CenterDot;Q=\underset{j=1}{\overset{D}{\mathrm{\&Sigma;}}}{z}_{\mathrm{ij}}\&times;q_j$ ，Z _i∈Z

Step 4: according to the category distribution situation of this K image, determine the classification of image to be classified.

The present invention is directed to these characteristics of froth images, froth images is abstracted into text, then just can as processing text, carry out Classification and Identification at semantic level to froth images.Owing to take full advantage of the local message of image, the froth images piecemeal is processed, and carry out twice abstract processing, obtain the description of froth images semantic level, one aspect of the present invention has improved the accuracy rate of classification, also solved on the other hand the semantic gap problem, identifying for froth images classification and operating mode provides a kind of new thinking.After the froth images classification, just identified current working, be used for thus instructing, optimize and produce, can reduce dosing, improve concentrate grade and mineral recovery rate.

[description of drawings]

Fig. 1 is the froth images that belongs to different vocabulary;

Fig. 2 is that the word bag of froth images is described synoptic diagram.

[embodiment]

Below in conjunction with accompanying drawing the specific embodiment of the present invention is described, according to different operating modes froth images is divided into 4 classes, the length of foam state vocabulary is chosen for 8.

1. the foam state vocabulary based on foam textural characteristics and color characteristic generates

From a collection of froth images of certain factory's floatation process collection in worksite totally 400 width of cloth.For this 400 image, it is 960 * 960 pixel sizes that every width of cloth image is all intercepted, carry out pre-service after, evenly be divided into 10 * 10.To each little piecemeal, extract its textural characteristics parameter and color parameter, obtain the low-level image feature vector of one 1 * 6 dimension.Therefore obtain altogether 40000 vectors, and these 40000 vectors are carried out K mean cluster (initial cluster center is selected at random), obtain D center (because the foam state vocabulary length of choosing is 8, D=8) here.Table 1 is resulting value when choosing 8 cluster centres, has formed the foam state vocabulary that contains 8 vocabulary.Fig. 1 has provided the froth images sub-block corresponding to different 8 vocabulary.

Table 1 contains the foam state vocabulary of 8 vocabulary

Foam state vocabulary	Foam state vocabulary vector
		Foam state vocabulary 1	(0.4597,1.5542,0.2826,0.3183,0.9159,1.0807)
Foam state vocabulary 2	(0.2767,2.0477,0.4839,0.2399,0.8663,1.0650)
		Foam state vocabulary 3	(0.2636,2.1293,0.5499,0.3051,0.8526,1.0782)
Foam state vocabulary 4	(0.4028,1.8784,0.5426,0.3011,0.8675,1.0853)
		Foam state vocabulary 5	(0.5218,1.5766,0.6347,0.2590,0.8817,1.0551)
Foam state vocabulary 6	(0.6718,1.1333,0.4545,0.3190,0.9211,1.0424)
		Foam state vocabulary 7	(0.6566,1.0234,0.3362,0.4186,0.9394,1.0407)
Foam state vocabulary 8	(0.3485,1.5968,0.1562,0.6743,0.9349,1.0704)

2. the word bag of training set froth images is described

The industry spot history image of different operating modes for 4 class correspondences of expertise artificial division, every class is got 100 width of cloth image, i.e. M=100.At first, with every width of cloth image interception to 960 * 960 pixel sizes, and evenly be divided into 10 * 10; To each little piecemeal, try to achieve its textural characteristics parameter and color parameter, obtain the low-level image feature vector of one 1 * 6 dimension; The low-level image feature of each little piecemeal vector is compared (calculating Euclidean distance) with the foam state vocabulary in the foam state vocabulary, it to the foam state vocabulary in which vocabulary the most similar (Euclidean distance is the shortest), just this piecemeal is designated as that vocabulary; At last, utilize the method for word bag, obtain the vector set C of every class image ₁, C ₂, C ₃, C ₄, C wherein _i=[c _I1c _I2C _IM] ^T, c _IjRepresenting the vector representation of j image of i class, is the row vector of 1 * D.

Fig. 2 has provided the word bag of a certain width of cloth image and has described process.Among Fig. 2,1. being image behind the piecemeal, 2. is foam state vocabulary vector table, 3. is that word bag corresponding to image described.Will be 1. behind the piecemeal, 2. the contrast vocabulary can obtain the corresponding vocabulary of each piecemeal, statistics 1. in the frequency that occurs of each vocabulary, 3. the word bag that obtains froth images is described.

3. carry out the froth images classification

To the froth images t to be sorted of real-time acquisition, the method for word bag is described, obtains vectorial Q _tAfter, just can classify to it as follows.

(1) inner product of vectors method: calculate respectively Q _tWith C ₁, C ₂, C ₃, C ₄Each vectorial inner product sum in the set, maximum with which kind of inner product sum, just which kind of image i is divided into.Way is as follows:

Word bag vector representation Q for the froth images to be sorted of trying to achieve _t=[2,76,16,4,0,2,0,0]; Each classification C in the training set ₁, C ₂, C ₃, C ₄The value of vector set is as follows respectively:

	First kind vector value (i=1)	Equations of The Second Kind vector value (i=2)	The 3rd class vector value (i=3)	The 4th class vector value (i=4)
					c i1	[2,64,22,4,2,2,4,0]	[0,4,2,3,0,2,82,7]	[2,0,2,4,15,65,12,0]	[32,0,2,4,0,0,2,60]
c i2	[4,80,10,0,0,0,3,3]	[1,4,2,4,0,2,72,15]	[2,4,0,12,70,10,2,0]	[25,1,2,4,0,0,12,56]
					c i3	[0,68,21,1,4,2,1,3]	[2,4,2,0,4,11,62,15]	[1,10,2,1,45,40,1,0]	[52,10,2,0,1,0,2,33]
c i4	[8,62,20,4,4,2,0,0]	[1,5,2,4,0,2,74,12]	[2,0,12,0,62,20,4,0]	[23,0,2,4,1,5,10,55]
					c i5	[5,72,15,0,0,4,2,2]	[0,0,5,2,2,1,68,22]	[0,0,10,3,80,6,0,1]	[42,1,2,6,0,0,12,37]
c i6	[4,76,11,4,5,0,0,0]	[1,2,8,2,0,0,77,10]	[9,0,0,2,72,12,0,5]	[34,10,2,4,0,0,0,50]
					c i7	[4,70,26,0,0,0,0,0]	[6,0,3,3,0,12,66,10]	[0,0,11,5,22,60,2]	[56,0,0,4,0,0,12,28]
...	...	..	...	...
					c i100	[2,76,18,2,0,0,1,1]	[4,4,3,0,0,9,67,13].	[2,4,2,0,35,45,11,1]	[45,10,2,0,0,1,2,40]

Calculate Q _tWith each vectorial inner product in each class and.Maximum with the inner product of which classification, just with Q _tWhich kind of is classified as.By formula ${\mathrm{SC}}_i=\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}\mathrm{Sim}(c_{\mathrm{ij}},Q_t),i=\mathrm{1,2,3,4}$ With $k=\underset{i^{*}}{\arg \max }{\mathrm{SC}}_i,i=\mathrm{1,2,3,4}$ , M=100 calculates SC here ₁=532040, SC ₂=23450, SC ₃=4324, SC ₄=5480.So with Q _tBe classified as the First Kind Graph picture.

(2) k nearest neighbor method: from training set C ₁, C ₂, C ₃, C ₄In select capable vector and Q _tA most similar K image is judged the classification under the new images of real-time acquisition according to the classification under K the image.

Table 2 and table 3 are respectively inner product of vectors method and the classification results of k nearest neighbor method when choosing different foam state vocabulary length (D=8,12,16) and different images piecemeal number (m=8,10,12).Data can find out that the piecemeal number of image and the scale of foam state vocabulary are influential to classification results from table.The more classifying qualities of foam state vocabulary are better, but the piecemeal of image is not The more the better.If the piecemeal number is (being that each piecemeal is too small) too much, because there is noise pollution (spot zone etc.) in image, the textural characteristics that obtains and color characteristic can not characterize its actual value, will cause vision word matched mistake; And the foam state vocabulary is larger, and calculated amount is also larger.Therefore, need to consider real-time and accuracy, choose the foam state vocabulary of suitable length, and carry out rational image block, generally get m=5 ~ 20.

Table 2 is based on the inner product of vectors classification results of VSM

Table 3 is based on the k nearest neighbor classification results of VSM

Conclusion: froth images sorting technique of the present invention can be carried out Classification and Identification to froth images more exactly, thereby determine current production belong to excellent, good, in, poor which kind of operating mode, be used for thus instructing, optimizing and produce, effectively reduce additive amount of medicament, improved concentrate grade and mineral recovery rate.Bubble sort method of the present invention is effective.

Claims (3)

1. sorting technique based on the mineral floating froth images is characterized in that comprising following process:

(1) generation is based on the floatation foam image foam state vocabulary of textural characteristics and color characteristic

Utilize the gray level co-occurrence matrixes algorithm to extract the parametric texture of froth images, comprise angle second moment, entropy, contrast, unfavourable balance square and correlativity; Extract the color characteristic of froth images, i.e. relative red component:

Gray level co-occurrence matrixes is by the joint probability density P(i between the image gray levels, j, d, θ) matrix that consists of, its definition direction is that θ, spacing are the gray level co-occurrence matrixes P(i of d, j, d, θ) and be the value of the capable i column element of i of co-occurrence matrix, θ gets 0 ⁰, 45 ⁰, 90 ⁰With 135 ⁰4 directions are established the digital picture that I (x, y) is a width of cloth two dimension, and its size is U * V pixel, and x, y are respectively the pixel coordinate value of pixel, then for different θ, P(i, j, d, θ) computing method are as follows:

P(i,j,d,0°)=Num{(x ₁,y ₁)(x ₂,y ₂)∈U*V|x ₁-x ₂=0,

|y ₁-y ₂|=d;I(x ₁,y ₁)=i,I(x ₂,y ₂)=j}

P (i, j, d, 45 °)=Num{ (x ₁, y ₁) (x ₂, y ₂) ∈ U*V| (x ₁-x ₂=d, y ₁-y ₂=d) or

(x ₁-x ₂=-d,y ₁-y ₂=-d);I(x ₁,y ₁)=i,I(x ₂,y ₂)=j}

P(i,j,d,90°)=Num{(x ₁,y ₁)(x ₂,y ₂)∈U*V|x ₁-x ₂=d,

y ₁-y ₂=0;I(x ₁,y ₁)=i,I(x ₂,y ₂)=j}

P (i, j, d, 135 °)=Num{ (x ₁, y ₁) (x ₂, y ₂) ∈ U*V| (x ₁-x ₂=d, y ₁-y ₂=-d) or

(x ₁-x ₂=-d,y ₁-y ₂=d);I(x ₁,y ₁)=i,I(x ₂,y ₂)=j}

Wherein, Num{X} represents to gather the element number among the X;

Texture, color parameter computing formula and be described below:

The angle second moment

$\mathrm{ASM}=\underset{i,j}{\mathrm{\&Sigma;}}{\left\{P(i,j|d,\mathrm{\&theta;})\right\}}^2$

Entropy

$\mathrm{IDM}$

Contrast

$\mathrm{CON}=\mathrm{\&Sigma;}{(i-j)}^{2}P(i,j|d,\mathrm{\&theta;})$

The unfavourable balance square

$\mathrm{IDM}=\underset{i,j}{\mathrm{\&Sigma;}}P(i,j|d,\mathrm{\&theta;})/[1+{(i-j)}^2]$

Correlativity

$\mathrm{COR}=\underset{i,j}{\mathrm{\&Sigma;}}(i-{\mathrm{\&mu;}}_x)(j-{\mathrm{\&mu;}}_y)P(i,j|d,\mathrm{\&theta;})/{\mathrm{\&sigma;}}_x{\mathrm{\&sigma;}}_y$

${\mathrm{\&mu;}}_x=\underset{i}{\mathrm{\&Sigma;}}i\underset{j}{\mathrm{\&Sigma;}}p(i,j|d,\mathrm{\&theta;})$ ， ${\mathrm{\&mu;}}_y=\underset{j}{\mathrm{\&Sigma;}}j\underset{i}{\mathrm{\&Sigma;}}p(i,j|d,\mathrm{\&theta;})$

${\mathrm{\&sigma;}}_x=\underset{i}{\mathrm{\&Sigma;}}{(i-{\mathrm{\&mu;}}_x)}^2\underset{j}{\mathrm{\&Sigma;}}p(i,j|d,\mathrm{\&theta;})$ ， ${\mathrm{\&sigma;}}_y=\underset{j}{\mathrm{\&Sigma;}}{(j-{\mathrm{\&mu;}}_y)}^2\underset{i}{\mathrm{\&Sigma;}}p(i,j|d,\mathrm{\&theta;})$

Relative red component

$R_{\mathrm{relative}}=\frac{R_{\mathrm{red}}}{R_{\mathrm{gray}}}$

In the formula, R _RedAnd R _GrayRepresent respectively red component average and gray average;

Froth images foam state vocabulary generative process is as follows:

Step 1: from image library, choose N width of cloth image, all image category of covering that N width of cloth image will be wide as far as possible, every width of cloth image all is truncated to a certain identical pixel size L _x* L _y

Step 2: for the N width of cloth image after the intercepting, every width of cloth image evenly is divided into m * m piece,

Step 3: for each piecemeal, obtain its textural characteristics value and relative red color component value, consist of the low-level image feature vector description of one 1 * 6 dimension;

Step 4: all low-level image feature vector descriptions are carried out the K-means cluster, and the D that an obtains cluster centre is the foam state vocabulary;

(2) the word bag of froth images is described

After obtaining the foam state vocabulary of froth images, method that just can the word bag is described each width of cloth froth images, obtains a vector representation of image.

The word bag of froth images is described, and detailed process is as follows:

Step 1: the image pixel intercepting is L _x* L _yPixel size;

Step 2: image evenly is divided into m * m piece;

Step 3: for each piecemeal, obtain its textural characteristics and relative red color component value, consist of one 1 * 6 dimension low-level image feature vector description;

Step 4: by calculating the Euclidean distance of each foam state vocabulary in this low-level image feature vector description and the foam state vocabulary, measure the similarity of all vocabulary in the vector description of each piecemeal and the foam state vocabulary, the most similar to which foam state vocabulary, which vocabulary just it is demarcated is;

Step 5: add up each foam state vocabulary frequency of occurrence, obtain the word bag vector representation of image;

(3) utilize vector space model to carry out the froth images classification

Inner product of vectors method or k nearest neighbor method are trained froth images classification and are classified, and obtain the different classes of of image.

2. the sorting technique based on the mineral floating froth images according to claim 1, it is characterized in that: in (1) or (2) described step 2, the m value is 5 ~ 20.

3. the sorting technique based on the mineral floating froth images according to claim 1, it is characterized in that: the process of the described inner product of vectors method of step (3) is:

For the image in each classification in the training set, the method for word bag is described, and obtains such word bag vector set C _i, i=1,2 ..., N _S, N _SClassification number for classified image in the training set; When getting access to new realtime graphic, the method for this image word bag is described, obtain the vector representation of image, gather C by calculating this word bag vector with the word bag vector of all kinds of images _iSimilarity classify;

Step 1: will train the image collection word bag model of each classification in the picture library to be described the vector set that obtains such image, each classification image is got M width of cloth image, M value 10 ~ 200, and then the word bag vector set of i class image is combined into

$C_i={\left[\begin{array}{cccc}{c}_{i1}& c_{i2}& \&CenterDot;\&CenterDot;\&CenterDot;& c_{\mathrm{iM}}\end{array}\right]}^T,i=\mathrm{1,2},...,N_S$

In the formula, c _IjThe word bag that represents j image of i class is described, and is the row vector of 1 * D;

Step 2: for image to be sorted, obtain its word bag and describe Q, Q=[q ₁q ₂Q _D] be the row vector of 1 * D;

Step 3: with Q and C _iThe vectorial c of in the class each _IjCarry out similarity and calculate, measure its similarity with the inner product of vector, obtain Q and each classification set C _iThe similarity sum SC of middle institute directed quantity _i

${\mathrm{SC}}_i=\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}\mathrm{Sim}(c_{\mathrm{ij}},Q),i=\mathrm{1,2,3}...,N_S$

In the formula, $\mathrm{Sim}(W,Q)=W\&CenterDot;Q=\underset{i=1}{\overset{D}{\mathrm{\&Sigma;}}}{w}_i\&times;q_i$ , W=[w ₁w ₂W _D] be the row vector of 1 * D;

Step 4: Q is classified as the k class of similarity maximum,

Images