CN101819410A - Intelligent spectrophotometry artificial neural network training method for testing water quality - Google Patents

Abstract

The invention relates to an artificial neural network training method for testing water quality. In the method, training data can be automatically generated and the artificial neural network can be automatically trained. The technical scheme is characterized by comprising the following steps of: reading the histogram data of three primary colors of red (R), green (G) and blue (B) of digital images; mapping the histogram data to two-dimensional planes, wherein the mapping result is plane PR(x, y), plane PG(x, y) and plane PB(x, y); extracting 7-dimensional characteristic vectors from the plane PR(x, y), the plane PG(x, y) and the plane PB(x, y) respectively, wherein seed characteristic vectors with the total number M correspond to the content concentration of a certain substance in the tested water sample; and training the artificial neural network according to the M seed characteristic vectors {VRGBi, wherein i is equal to 1, 2,..., and M}.

Description

The intelligent spectrophotometric artificial neural network training method that is used for water quality detection

Technical field

The invention belongs to water quality detection intelligence spectrophotometric artificial neural network training method field, especially a kind ofly can generate training data automatically, train artificial neural network's the artificial neural network's training method that is used for water quality detection automatically.

Background technology

Artificial neural network's training method is used for color recognition, though report for work in the research abroad, but its training result is not accurate enough, the data sampling mode lacks the ability of meticulous depiction color variables, training effect is not good, and convergence is slow, resultant error is big, can not perform well in the automatic colorimetric of spectrophotometric method.

Summary of the invention

The purpose of this invention is to provide and a kind ofly can generate training data automatically, train artificial neural network's the artificial neural network's training method that is used for water quality detection automatically.

Technical scheme of the present invention is: the intelligent spectrophotometric artificial neural network training method that is used for water quality detection, it is characterized in that comprising the following steps: to read the histogram data of digital picture R, G, B three primary colours, histogram data is mapped to two dimensional surface, promptly tie up variable as transverse axis with histogrammic first, unit is the quantization unit value of image primary colours, tie up variable as the longitudinal axis with histogrammic second, unit is through the highest numerical value of histogram after the normalization, mapping result is plane P R (x, y), plane PG (x, y) and plane PB (x, y)

To described each plane P R (x, y), (x, y) (x y) extracts 7 dimensional feature vectors, i.e. V respectively to plane PG with plane PB ^R=(V ^R ₁, V ^R ₂, V ^R ₃... V ^R ₇), V ^G=(V ^G ₁, V ^G ₂, V ^G ₃... V ^G ₇), V ^B=(V ^B ₁, V ^B ₂, V ^B ₃... V ^B ₇), V ^RProper vector corresponding flat PR (x, y), V ^GProper vector corresponding flat PG (x, y), V ^BProper vector corresponding flat PB (x, y), each 7 dimensional feature vector is defined as follows, N=7 in the formula:

${V^R}_1=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}\mathrm{PR}(x,y);\·\·\·\left(1\right)$

${V^R}_2=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}\mathrm{xPR}(x,y)/V_1;\·\·\·\left(2\right)$

${V^R}_3=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}\mathrm{yPR}(x,y)/V_1;\·\·\·\left(3\right)$

${V^R}_4=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}2\mathrm{xyPR}(x,y)/V_1;\·\·\·\left(4\right)$

${V^R}_5=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}{x}^2\mathrm{PR}(x,y)/V_1;\·\·\·\left(5\right)$

${V^R}_6=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}{y}^2\mathrm{PR}(x,y)/V_1;\·\·\·\left(6\right)$

V ^R ₇＝V ₄/(V ₅-V ₆)；...(7)

(x is y) with plane PB (x, y) the proper vector V on for two primary color plane PG of other G and B ^G, V ^BWith defining with quadrat method;

M is training artificial neural network's seed characteristics vector total number in the above-mentioned formula, and the span of M is determined (M is a number of samples) by the number of samples in the chemical reagent colourimetry, and total number is that M seed characteristics vector representation is V ^RGB ₁=(V ^R ₁, V ^G ₁, V ^B ₁), V ^RGB ₂=(V ^R ₂, V ^G ₂, V ^B ₂) ... V ^RGB _M=(V ^R _M, V ^G _M, V ^B _M); Be designated as { V ^RGB _i| i=1,2 ... M), the concentration of certain content of material in M corresponding one by one detected water sample of proper vector, and according to the increase ordering of concentration, i.e. the 1st seed characteristics vector V ^RGB ₁Represent that certain content of material is minimum, the 2nd proper vector V ^RGB ₂Represent that certain content of material is inferior low, and the like, M V ^RGB _MRepresent that certain content of material is the highest, M proper vector be expressed as plane P R (x, y), plane PG (x, y) and plane (x, M point in the N-dimension space that maps out on y);

According to M seed characteristics vector { V ^RGB _i| i=1,2 ... M} trains the artificial neural network, comprises the following steps:

(1) extract per two proper vectors successively from the seed characteristics vector of M, form neighbours' " vector to ", this vector is to being designated as (V ^{RGB, i}, V ^{RGB, j}), i=1 here, 2 ... M-1; J=i+1;

(2) choose primary color space R, in this primary color space, neighbours' " vector to " (V ^{RGB, i}, V ^{RGB, j}) become (V ^{R, i}, V ^{R, j});

(3) with (V ^{R, i}, V ^{R, j}) successively to the projection of N-dimension space, on each projector space, obtain projection value, calculate the midrange on each projector space, serve as basis definition upper limit value and lower limit value with this midrange, higher limit V ^{R, ij} _{K, MAX}Be defined as the numerical value after midrange increases by 10% increment, lower limit V ^{R, ij} _{K, MIN}Be defined as the numerical value after midrange reduces by 10% increment;

Thereby on the 1st dimension space, V is arranged ^{R, ij} _{1, MID}, V ^{R, ij} _{1, MIN}, V ^{R, ij} _{1, MAX}On the 2nd dimension space, V is arranged ^{R, ij} _{2, MID}, V ^{R, ij} _{2, MIN}, V ^{R, ij} _{2, MAX}And the like, on the whole N-dimension space { V is arranged ^{R, ij} _{K, MID}, V ^{R, ij} _{K, MIN}, V ^{R, ij} _{K, MAX}| k=1,2 ..., M}; So on the N-dimension space, form the data set G of gang at each neighbour " vector to " ^R _Ij, G here ^R _Ij={ V ^{R, ij} _{K, MID}, V ^{R, ij} _{K, MIN}, V ^{R, ij} _{K, MAX}| k=1,2 ..., M};

(4) finish to all i j neighbours' " vector to " (V ^{RGB, i}, V ^{RGB, j}) midrange and the calculating of upper limit value and lower limit value, i=1,2 ... M-1; J=i+1; Thereby obtain (x, y) the training data group G of the artificial neural network on the plane on this basis at PR ^R _SUM={ G ^R _Ij| i=1,2 ... M-1; J=i+1};

(5) choose primary color space G and B, finish same steps as (3), (4), thus obtain PG (x, y), PB (x, y) the training data group G of the artificial neural network on the plane ^G _SUM={ G ^G _Ij| i=1,2 ... M-1; J=i+1} and G ^B _SUM={ G ^B _Ij| i=1,2 ... M-1; J=i+1}, so far data generate and finish, and finish the training to the artificial neural network thus.

Effect of the present invention is: the present invention is a kind of to the training data generation technique of artificial neural network in intelligent spectrophotometric is used.Present technique is according to the N dimensional feature vector that extracts from Digital Image Processing, generation is to artificial neural network's training matrix data set, can the implementation pattern recognition function through the artificial neural network of this training, finish automatic colorimetric detection, under spectrophotometric, recognize contaminants associated in the water sample water sample.This training data generation technique and artificial neural network are applicable to that portable broad spectrum spectroscopic luminosity liquid trace element detects the wireless sensing instrument, be convenient on embedded system, realize, has automatic generation training data, automatically train the artificial neural network, proofread and correct the characteristics of portable detector fast, be suitable for wide-spectrum water and detect application.

The present invention is described further below in conjunction with drawings and Examples.

Description of drawings

Fig. 1 is a procedure chart of the present invention;

Fig. 2 is a schematic diagram of the present invention.

Embodiment

Present technique is made up of three parts: one, from lower dimensional space, and through the Compound Mappings method, definition N dimensional feature vector.At first by histogram to the two-dimensional geometry space conversion, and the N dimension space that reaches proper vector is on this basis realized.Two, calculate the midrange of any two proper vectors behind the k-th space projection, and the border higher limit and the lower limit in k-th space be set hereinto on the basis of point value, with k-th space midrange is to form k dimensional feature data field between the higher limit of reference and the lower limit, gathers as the training data point on the k-th space.Three, the method for fetching data on the k-th space is extended in the N dimension space, form one group comprise RGB three primary colours plane, to artificial neural network's training matrix data set, thereby finish the work of automatic generation training data.

Arthmetic statement of the present invention is shown in figure one.A is the first step of algorithm among the figure one, promptly obtains the histogram of digital picture R, G, B three primary colours, and the histogram of three primary colours is shown in figure two.Algorithm B disposes for judging whether each histogram among the figure one, if dispose, has then finished the generation work of training data, enters C, the computing that end data generates; Otherwise enter D, histogram be mapped to the 2-dimensional plane, mapping result be PR (x, y), PG (x, y) and PB (x, y); Be three 2-dimensional planes.

The method that histogram data is mapped to two dimensional surface is: as transverse axis, unit is the quantization unit value (quantization level) of image primary colours with the histogrammic first dimension variable.As primary colours is 8 bits, and then the first dimension variable transverse axis has 256 values, i.e. 28 powers.As the longitudinal axis, unit is through the highest numerical value of histogram after the normalization with the histogrammic second dimension variable.As image analytic degree is NxN, and then the highest numerical value of histogram is N ², it is made its normalization divided by N, then the highest numerical value of histogram after the normalization is N.The two dimensional surface of Xing Chenging contains the histogram after the normalization like this, becomes the result after the two dimensional surface mapping.

On the basis of this 2-dimension space, to each plane P R (x, y), PG (x, y) and PB (x y) extracts N dimensional feature vector (N=7) respectively, shown in E among the figure, V is arranged promptly ^R=(V ^R ₁, V ^R ₂, V ^R ₃... V ^R ₇) proper vector corresponding PR (x, y) plane, V ^G=(V ^G ₁, V ^G ₂, V ^G ₃... V ^G ₇) proper vector corresponding PG (x, y) plane, V ^B=(V ^B ₁, V ^B ₂, V ^B ₃... V ^B ₇) the corresponding PB of proper vector (x, y) plane.Each N-dimensional feature vector is defined as follows:

${V^R}_1=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}\mathrm{PR}(x,y);\·\·\·\left(1\right)$

${V^R}_2=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}\mathrm{xPR}(x,y)/V_1;\·\·\·\left(2\right)$

${V^R}_3=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}\mathrm{yPR}(x,y)/V_1;\·\·\·\left(3\right)$

${V^R}_4=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}2\mathrm{xyPR}(x,y)/V_1;\·\·\·\left(4\right)$

${V^R}_5=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}{x}^2\mathrm{PR}(x,y)/V_1;\·\·\·\left(5\right)$

${V^R}_6=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}{y}^2\mathrm{PR}(x,y)/V_1;\·\·\·\left(6\right)$

V ^R ₇＝V ₄/(V ₅-V ₆)；...(7)

Two primary color plane PG of other G and B (x, y) and PB (x, y) the proper vector V on ^G, V ^BWith defining with quadrat method.

Training artificial neural network's seed characteristics vector total number is M, is expressed as V ^RGB ₁=(V ^R ₁, V ^G ₁, V ^B ₁), V ^RGB ₂=(V ^R ₂, V ^G ₂, V ^B ₂) ... V ^RGB _M=(V ^R _M, V ^G _M, V ^B _M); Be designated as { V ^RGB _i| i=1,2 ... M}, the concentration of certain content of material in the corresponding one by one detected water sample of this M proper vector, and according to the increase of concentration ordering, also, the 1st seed characteristics vector V ^RGB ₁Represent that certain content of material is minimum, the 2nd proper vector V ^RGB ₂Represent that certain content of material is inferior low, and the like, so M V ^RGB _MRepresent that certain content of material is the highest.Can M proper vector be expressed as each plane P R (x, y), PG (x, y) and PB (x, M point in the N-dimension space that maps out on y).

According to M seed characteristics vector { V ^RGB _i| i=1,2 ... M}, we set up artificial neural network's data creation method, are used for the training to the artificial neural network.

Step 1 is extracted per two proper vectors successively from total number is the seed characteristics vector of M, form neighbours' " vector to ", and promptly among the figure one shown in the F, this vector is to being designated as (V ^{RGB, i}, V ^{RGB, j}), i=1 here, 2 ... M-1; J=i+1;

Step 2 is chosen primary color space R, in this primary color space, and neighbours' " vector to " (V ^{RGB, i}, V ^{RGB, j}) become (V ^{R, i}, V ^{R, j});

Step 3, on this basis, with (V ^{R, i}, V ^{R, j}) successively to the projection of N-dimension space, on each projector space, obtain projection value, shown in G among the figure one; And the midrange of calculating on each projector space.Midrange is a basis definition upper limit value and lower limit value thus, shown in H among the figure one.Higher limit V ^{R, ij} _{K, MAX}After being defined as the increment of midrange increase by 10%, near V ^{R, j} _kNumerical value, lower limit V ^{R, ij} _{K, MIN}After being defined as the increment of midrange minimizing 10%, near V ^{R, i} _kNumerical value.

Thereby on the 1st dimension space, have: V ^{R, ij} _{1, MID}, V ^{R, ij} _{1, MIN}, V ^{R, ij} _{1, MAX}On the 2nd dimension space, have: V ^{R, ij} _{2, MID}, V ^{R, ij} _{2, MIN}, V ^{R, ij} _{2, MAX}And the like, on the whole N-dimension space { V is arranged ^{R, ij} _{K, MID}, V ^{R, ij} _{K, MIN}, V ^{R, ij} _{K, MAX}| k=1,2 ..., M}; So on the N-dimension space, form the data set G of gang at each neighbour " vector to " ^R _Ij, G here ^R _Ij={ V ^{R, ij} _{K, MID}, V ^{R, ij} _{K, MIN}, V ^{R, ij} _{K, MAX}| k=1,2 ..., M}.

Step 4 is finished all i, j neighbours' " vector to " (V ^{RGB, i}, V ^{RGB, j}) midrange and the calculating of upper limit value and lower limit value, i=1,2 ... M-1; J=i+1; Thereby obtain (x, y) the training data group G of the artificial neural network on the plane on this basis at PR ^R _SUM={ G ^R _Ij| i=1,2 ... M-1; J=i+1} is shown in I among the figure one.

Step 5 is chosen primary color space G and B, finishes same steps as three, four; Thereby obtain PG (x, y), PB (x, y) the training data group G of the artificial neural network on the plane ^G _SUM={ G ^G _Ij| i=1,2 ... M-1; J=i+1} and G ^B _SUM={ G ^B _Ij| i=1,2 ... M-1; J=i+1}, shown in J among the figure one, so far data generate and finish, and finish the training to the artificial neural network thus.

Claims (1)

1. the intelligent spectrophotometric artificial neural network training method that is used for water quality detection, it is characterized in that comprising the following steps: to read the histogram data of digital picture R, G, B three primary colours, histogram data is mapped to two dimensional surface, promptly tie up variable as transverse axis with histogrammic first, unit is the quantization unit value of image primary colours, tie up variable as the longitudinal axis with histogrammic second, unit is through the highest numerical value of histogram after the normalization, mapping result is plane P R (x, y), plane PG (x, y) and plane PB (x, y)

To described each plane P R (x, y), (x, y) (x y) extracts 7 dimensional feature vectors, i.e. V respectively to plane PG with plane PB ^R=(V ^R ₁, V ^R ₂, V ^R ₃... V ^R ₇), V ^G=(V ^G ₁, V ^G ₂, V ^G ₃... V ^G ₇), V ^B=(V ^B ₁, V ^B ₂, V ^B ₃... V ^B ₇), V ^RProper vector corresponding flat PR (x, y), V ^GProper vector corresponding flat PG (x, y), V ^BProper vector corresponding flat PB (x, y), each 7 dimensional feature vector is defined as follows, N=7 in the formula:

${V^R}_1=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}\mathrm{PR}(x,y);$

${V^R}_2=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}\mathrm{xPR}(x,y)/V_1;$

${V^R}_3=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}\mathrm{yPR}(x,y)/V_1;$

${V^R}_4=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}2\mathrm{xyPR}(x,y)/V_1;$

${V^R}_5=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}{x}^2\mathrm{PR}(x,y)/V_1;$

${V^R}_6=\underset{x=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{M-1}{\mathrm{\Σ}}}{y}^2\mathrm{PR}(x,y)/V_1;$

V ^R ₇＝V ₄/(V ₅-V ₆)；

(x is y) with plane PB (x, y) the proper vector V on for two primary color plane PG of other G and B ^G, V ^BWith defining with quadrat method;

M is training artificial neural network's seed characteristics vector total number in the above-mentioned formula, and the span of M is determined that by the number of samples in the chemical reagent colourimetry promptly M is a number of samples; Total number is that M seed characteristics vector representation is V ^RGB ₁=(V ^R ₁, V ^G ₁, V ^B ₁), V ^RGB ₂=(V ^R ₂, V ^G ₂, V ^B ₂) ... V ^RGB _M=(V ^R _M, V ^G _M, V ^B _M); Be designated as { V ^RGB _i| i=1,2 ... the concentration of certain content of material among the M}, M corresponding one by one detected water sample of proper vector, and according to the increase ordering of concentration, i.e. the 1st seed characteristics vector V ^RGB ₁Represent that certain content of material is minimum, the 2nd proper vector V ^RGB ₂Represent that certain content of material is inferior low, and the like, M V ^RGB _MRepresent that certain content of material is the highest, M proper vector be expressed as plane P R (x, y), plane PG (x, y) and plane (x, M point in the N-dimension space that maps out on y);

According to M seed characteristics vector { V ^RGB _i| i=1,2 ... M} trains the artificial neural network, comprises the following steps:

(1) extract per two proper vectors successively from the seed characteristics vector of M, form neighbours' " vector to ", this vector is to being designated as (V ^{RGB, i}, V ^{RGB, j}), i=1 here, 2 ... M-1; J=i+1;

(2) choose primary color space R, in this primary color space, neighbours' " vector to " (V ^{RGB, i}, V ^{RGB, j}) become (V ^{R, i}, V ^{R, j});

(3) with (V ^{R, i}, V ^{R, j}) successively to the projection of N-dimension space, on each projector space, obtain projection value, calculate the midrange on each projector space, serve as basis definition upper limit value and lower limit value with this midrange, higher limit V ^{R, ij} _{K, MAX}Be defined as the numerical value after midrange increases by 10% increment, lower limit V ^{R, ij} _{K, MIN}Be defined as the numerical value after midrange reduces by 10% increment;

Thereby on the 1st dimension space, V is arranged ^{R, ij} _{1, MID}, V ^{R, ij} _{1, MIN}, V ^{R, ij} _{1, MAX}On the 2nd dimension space, V is arranged ^{R, ij} _{2, MID}, V ^{R, ij} _{2, MIN}, V ^{R, ij} _{2, MAX}And the like, on the whole N-dimension space { V is arranged ^{R, ij} _{K, MID}, V ^{R, ij} _{K, MIN}, V ^{R, ij} _{K, MAX}| k=1,2 ..., M}; So on the N-dimension space, form the data set G of gang at each neighbour " vector to " ^R _Ij, G here ^R _Ij={ V ^{R, ij} _{K, MID}, V ^{R, ij} _{K, MIN}, V ^{R, ij} _{K, MAX}| k=1,2 ..., M};

(4) finish to all i j neighbours' " vector to " (V ^{RGB, i}, V ^{RGB, j}) midrange and the calculating of upper limit value and lower limit value, i=1,2 ... M-1; J=i+1; Thereby obtain (x, y) the training data group G of the artificial neural network on the plane on this basis at PR ^R _SUM={ G ^R _Ij| i=1,2 ... M-1; J=i+1};

(5) choose primary color space G and B, finish same steps as (3), (4), thus obtain PG (x, y), PB (x, y) the training data group G of the artificial neural network on the plane ^G _SUM={ G ^G _Ij| i=1,2 ... M-1; J=i+1} and G ^B _SUM={ G ^B _Ij| i=1,2 ... M-1; J=i+1}, so far data generate and finish, and finish the training to the artificial neural network thus.

Images