CN106951919A - A kind of flow monitoring implementation method based on confrontation generation network - Google Patents

Abstract

A kind of flow monitoring implementation method based on confrontation generation network, comprises the following steps：(1) current image preprocessing；(2) image classification is carried out based on confrontation generation network；(3) measurement of rate of flow：Image classification result is corresponded with flow velocity interval；(4) state analysis：Abnormal state signal is sent when monitoring that water velocity exceeds predetermined threshold value.Beneficial effects of the present invention are mainly manifested in：The dual training of maker and arbiter effectively combines the advantage of identification and generative nature sorting algorithm and realizes unsupervised learning, the synthetic water stream picture that the maker output of generation network will be resisted substantially increases robustness of the grader to noisy current image with true picture collectively as the input of arbiter, is classified according to current image and corresponds to the quick measure of the default achievable flow velocity in flow velocity interval and be easy to the Classification Management of a large amount of flow information of water.

Description

A kind of flow monitoring implementation method based on confrontation generation network

Technical field

The present invention relates to a kind of flow monitoring implementation method, and in particular to a kind of flow monitoring based on confrontation generation network Implementation method, belongs to area of pattern recognition.

Background technology

Water velocity monitoring can be used directly or indirectly in power station intelligent scheduling, hydrologic monitoring and hazard forecasting, be water conservancy The work institutes such as project planning design, flood-control and drought relief and irrigation production are required, and fast and accurately flow monitoring can significantly improve water conservancy The science and the foresight of drought and waterlogging of Project Scheduling.Flow monitoring based on image recognition possesses inexpensive, high-precision Advantage, its core technology is image recognition.The sorting algorithm of image recognition is broadly divided into identification and the class of generative nature two, its In, the former, need to just can be directly to without number of tags using some classification mechanisms without explicitly be modeled to True Data distribution According to being classified, but it is easily related to the falseness in data produce over-fitting, it is particularly true in Nonlinear Classifier；The latter is then Generation model is parameterized by training depth Boltzmann machine or feed forward type neutral net etc. and by training self-encoding encoder straight Connect and data distribution or geometric attribute are modeled, this kind of learning method based on reconstruct attempts study, and those remain input sample The feature of this full detail, however, generation sample may not correctly reflect True Data feature to cause based on these features hardly possible Correctly to be classified.

Sorting technique generally requires a large amount of label informations, however, with the continuous accumulation of current amount of images, being adopted to all Sample to current image to carry out manual identification one by one be unpractical, therefore, based on unsupervised learning or merely with a small amount of label The image classification of the semi-supervised learning of information has more actual application value than traditional sorting technique based on supervised learning.This Outside, field node suffers from the interference of the unfavorable factors such as rain, mist and light and causes sampled images to include a large amount of noises, significantly Exacerbate classification difficulty.Therefore, how to improve grader and unsupervised or half are based on to the robustness of noisy current image and realization The current image classification of supervision feature learning has far reaching significance.

The content of the invention

The present invention is directed to the limitation of above-mentioned prior art, proposes confrontation generation network (Generative Adversarial Networks, GAN) tested the speed applied to the current based on image classification, by resisting train a pair of generations Device and arbiter each advantage and realize unsupervised or semi-supervised learning to combine generating algorithm and distinguished number, and maker is defeated The composograph gone out is with actual water stream picture while effectively increasing sorting algorithm to noisy current figure as the input of arbiter The robustness of picture.

The technical solution adopted for the present invention to solve the technical problems is：

A kind of flow monitoring implementation method based on confrontation generation network, comprises the following steps：

Step 1, current image preprocessing：RGB current images are converted into gray-scale map and histogram equalization is sequentially passed through And random cropping, then the image intensity value after processing is deployed and connected by row, gained column vector constitutes to be sorted Current image data set eigenmatrix X=[x₁,x₂,…,x_n], wherein sample x_iIt is the column vector that dimension is d；

Step 2, image classification is carried out based on confrontation generation network：Train the model of a pair of confrontation：Arbiter D purpose is Correct differentiation True Data and generation data simultaneously carry out classification so as to maximize differentiation accuracy rate to True Data, and maker G Purpose is then the data mixed the spurious with the genuine of generation and minimizes arbiter D differentiation accuracy rate, both training are a binary Zero-sum game process；H represents entropy, and E represents desired value, and p represents probability, and data and g represent True Data collection and generation number respectively According to collection, prior distribution is mapped as generating sample by maker G for P (z) random noise vector z by x=G (z)；

2.1 calculate the entropy of classification y marginal probability in True Data collection data according to formula (1)：

Wherein, D represents arbiter, and p (y |) represents y marginal probability, and N represents True Data collection data sample Number；

2.2 experiences for calculating conditional entropies of the image x from True Data collection data according to formula (2) are estimated：

Wherein, K represents image category (namely flow velocity is interval) number, and p (y=k |) represents that image belongs to classification k side Edge probability；

The experience of 2.3 conditional entropies for carrying out self-generating data set g to image x according to the Monte Carlo sampling of formula (3) is estimated to carry out Approach：

Wherein, M represents the quantity of independent sample, noise vector zⁱ~P (z)；

2.4 calculate the entropy of classification y marginal probability in generation data set g according to formula (4)：

2.5 unsupervised learnings are modeled, object function such as formula (5) institute of the non-supervision image classification based on confrontation generation network Show：

Wherein, L_DIt is arbiter D loss function, L_GIt is maker G loss function；

2.6 semi-supervised learnings are modeled：It is above-mentioned without prison when known to the classification information of part current image in training set data Semisupervised classification problem can be further converted to by superintending and directing classification problem, now, and maker G loss function keeps constant in formula (5), That is, L^label _G=L_G, and arbiter D loss function is adjusted to：

Wherein, λ is balance parameters, label data collection data^label={ (x¹,y¹),…,(x^L,y^L), L represents label data Number, CE [y, p (y | x, D)] represent data^labelThe condition distribution p (y | x, D) of middle image x true classification y and prediction classification Cross entropy, specifically calculate as shown in formula (7)：

2.7 model optimizations are solved：Because maker G and arbiter D can be micro-, therefore can be according to object function (5) Or (6) utilize stochastic gradient descent method to L_DAnd L_GCarry out alternately training；

Step 3, measurement of rate of flow：The classification results of step 2 are corresponded with predefined flow velocity interval；

Step 4, state analysis：Abnormal state signal is sent when monitoring that water velocity exceeds predetermined threshold value.

The technical concept of the present invention：The arbiter of housebroken unsupervised confrontation generation network can be used as a probabilistic type point Class device, the grader possesses higher robustness due to being trained through composograph and true picture simultaneously, utilizes on this basis (the far smaller than sample size of training set) there is the current image of classification information to carry out semi-supervised learning and be greatly improved on a small quantity Classification accuracy.Current image is classified based on above-mentioned confrontation generation network, then by classification results and each flow velocity area Between correspond realizing the quick measure of water velocity.

Beneficial effects of the present invention：The dual training of maker and arbiter effectively combines identification and generative nature classification The advantage of algorithm simultaneously realizes unsupervised learning, will resist the synthetic water stream picture and true figure of the maker output for generating network As the input collectively as arbiter substantially increases robustness of the grader to noisy current image, carried out according to current image Classify and correspond to default flow velocity interval and the quick measure of flow velocity and the Classification Management of a large amount of flow information of water can be achieved.This method Cost is relatively low, accuracy rate is higher, it is easy to promote.

Brief description of the drawings

Fig. 1 is the inventive method flow chart.

Fig. 2 a are current gradation of image figures.

Fig. 2 b are the current images after histogram equalization.

Fig. 3 a, Fig. 3 b, Fig. 3 c are flow velocity interval 0-0.25m/s, 0.25-0.5m/s, 0.5-0.75m/s current figure respectively As example.

Embodiment

The invention will be further described below.

A kind of flow monitoring implementation method based on confrontation generation network, comprises the following steps：

Step 1, current image preprocessing：RGB current images are converted into gray-scale map and histogram equalization is sequentially passed through And random cropping, then the image intensity value after processing is deployed and connected by row, gained column vector constitutes to be sorted Current image data set eigenmatrix X=[x₁,x₂,…,x_n], wherein sample x_iIt is the column vector that dimension is d；

Step 2, image classification is carried out based on confrontation generation network：Train the model of a pair of confrontation：Arbiter D purpose is Correct differentiation True Data and generation data simultaneously carry out classification so as to maximize differentiation accuracy rate to True Data, and maker G Purpose is then the data mixed the spurious with the genuine of generation and minimizes arbiter D differentiation accuracy rate, both training are a binary Zero-sum game process.H represents entropy, and E represents desired value, and data and g represent True Data collection and generation data set, maker respectively Prior distribution is mapped as generating sample by G for P (z) random noise vector z by x=G (z)；

2.1 calculate the entropy of classification y marginal probability in True Data collection data according to formula (1)：

Wherein, D represents arbiter, and p (y |) represents y marginal probability, and N represents True Data collection data sample Number；

2.2 experiences for calculating conditional entropies of the image x from True Data collection data according to formula (2) are estimated：

Wherein, K represents image category (namely flow velocity is interval) number, and p (y=k |) represents that image belongs to classification k side Edge probability；

The experience of 2.3 conditional entropies for carrying out self-generating data set g to image x according to the Monte Carlo sampling of formula (3) is estimated to carry out Approach：

Wherein, M represents the quantity of independent sample, noise vector zⁱ~P (z)；

2.4 calculate the entropy of classification y marginal probability in generation data set g according to formula (4)：

2.5 unsupervised learnings are modeled, object function such as formula (5) institute of the non-supervision image classification based on confrontation generation network Show：

Wherein, L_DIt is arbiter D loss function, L_GIt is maker G loss function；

2.6 semi-supervised learnings are modeled：It is above-mentioned without prison when known to the classification information of part current image in training set data Semisupervised classification problem can be further converted to by superintending and directing classification problem, now, and maker G loss function keeps constant in formula (5), That is, L^label _G=L_G, and arbiter D loss function is adjusted to：

Wherein, λ is balance parameters, label data collection data^label={ (x¹,y¹),…,(x^L,y^L), L represents label data Number, CE [y, p (y | x, D)] represent data^labelThe condition distribution p (y | x, D) of middle image x true classification y and prediction classification Cross entropy, specifically calculate as shown in formula (7)：

2.7 model optimizations are solved：Because maker G and arbiter D can be micro-, therefore can be according to object function (5) Or (6) utilize stochastic gradient descent method to L_DAnd L_GCarry out alternately training；

Step 3, measurement of rate of flow：The classification results of step 2 are corresponded with predefined flow velocity interval；

Step 4, state analysis：Abnormal state signal is sent when monitoring that water velocity exceeds predetermined threshold value.Example：

Picture is pre-processed：Because monitoring point is in outdoor, to the shooting of current image unavoidably by weather (such as：Rain, snow and Mist) and the influence of the factor such as illumination variation.To weaken influence of these factors to picture quality, the RGB of current image is schemed to convert Gone forward side by side column hisgram equalization processing for gray-scale map, enhancing contrast is so that water wave profile becomes obvious, Fig. 2 (a) and 2 (b) It is gray-scale map and the figure through histogram equalization processing respectively.10 flow velocity intervals are predefined according to the historical data of the monitoring point (interval number can increase and decrease according to the requirement in practical application to precision), is 0-0.25m/s, 0.25-0.5m/s respectively, 0.5-0.75m/s, 0.75-1.0m/s, 1.0-1.25m/s, 1.25-1.5m/s, 1.5-2.0m/s, 2.0-2.5m/s, 2.5- 3.0m/s, 3m/s and more than, each flow velocity interval includes 30 current images (being used for semi-supervised learning) for having a class label, stream Fast threshold value is 2.3m/s.Fig. 3 (a)-(c) respectively show the current image in interval 1,2,3, and every current artwork size is 1000 ×750.The computation complexity of network is generated so as to effectively improve operation efficiency, by each Zhang Yuantu random croppings for reduction confrontation For the small size picture of 100 32 × 32.

The framework of confrontation generation network sets as follows：Convolutional neural networks and deconvolution are respectively adopted in arbiter and maker Neutral net, it is specific as shown in table 1, wherein, conv represents convolutional layer, and LeakyReLU represents that Leaky corrects linear unit, fc Full articulamentum is represented, stride represents maximum pondization operation max-pool strides；

The framework details of table 1

The parameter setting of confrontation generation network is as follows：The standard deviation of the gaussian noise added is 0.05, noise vector z's Dimension is set to 128, leak rate=0.1, crowd normalized scale B=100, i.e. every time from True Data collection data (or marks Sign data set data^label) and maker G (z) in the sample number randomly selected be 100, learning rate is calculated according to Adam Method carries out Automatic adjusument and the upper limit is 0.001；

Model training：Because maker G and arbiter D can be micro-, therefore can be according to object function (5) or (6) profit With stochastic gradient descent method to L_DAnd L_GCarry out alternately training：First, fix arbiter D and optimize maker G, make arbiter D's Differentiate that accuracy rate is as low as possible, then, fixed maker G simultaneously optimizes arbiter D, so as to improve arbiter D differentiation accuracy rate. For balance balance maker G and arbiter D, set arbiter D often to update maker G after 5 times and just update 1 time to avoid maker There is pattern and caved in G.In addition, the unstability to overcome confrontation generation network, takes following two during model training Measure：First, all layers of use batch normalization (batch to arbiter D all layers and maker G in addition to last layer Normalization), this causes each layer of activation value bounded effectively to prevent the pattern switching of maker and improve differentiation The Generalization Capability of device, then, by the way that gaussian noise is added through criticizing the arbiter hidden layer of normalized to realize arbiter Regularization.

Content described in this specification embodiment is only enumerating to the way of realization of inventive concept, protection of the invention Scope is not construed as being only limitted to the concrete form that embodiment is stated, protection scope of the present invention is also and in art technology Personnel according to present inventive concept it is conceivable that equivalent technologies mean.

Claims (1)

1. a kind of flow monitoring implementation method based on confrontation generation network, comprises the following steps：

Step 1, current image preprocessing：By RGB current images be converted to gray-scale map and sequentially pass through histogram equalization and with Machine is cut, and then the image intensity value after processing is deployed and connected by row, gained column vector constitutes water to be sorted Eigenmatrix X=[the x of flow image data collection₁,x₂,…,x_n], wherein sample x_iIt is the column vector that dimension is d；

Step 2, image classification is carried out based on confrontation generation network：Train the model of a pair of confrontation：Arbiter D purpose is correct Differentiation True Data and generation data simultaneously carry out classification so as to maximize differentiation accuracy rate to True Data, and maker G mesh Be then the data mixed the spurious with the genuine of generation and minimize arbiter D differentiation accuracy rate, both training are a binary zero-sums Gambling process；H represents entropy, and E represents desired value, and p represents probability, and data and g represent True Data collection and generation data set respectively, Prior distribution is mapped as generating sample by maker G for P (z) random noise vector z by x=G (z)；

2.1 calculate the entropy of classification y marginal probability in True Data collection data according to formula (1)：

$H_{data}\[p\left(y\right|D)\]=H\[\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}p\left(y\right|x^i,D)\]---\left(1\right)$

Wherein, D represents arbiter, and p (y |) represents y marginal probability, and N represents True Data collection data number of samples；

2.2 experiences for calculating conditional entropies of the image x from True Data collection data according to formula (2) are estimated：

$\begin{array}{c}{E}_{x~data}\[H\[p\left(y\right|x,D)\]\]=\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}H\[p\left(y\right|x^i,D)\]\\ =\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}-\underset{k=1}{\overset{K}{\Σ}}p(y=k|x^i,D)\log p(y=k|x^i,D)\end{array}---\left(2\right)$

Wherein, K represents image category (namely flow velocity is interval) number, and the edge that p (y=k |) expression images belong to classification k is general Rate；

The experience estimation of 2.3 conditional entropies for carrying out self-generating data set g to image x according to the Monte Carlo sampling of formula (3) is approached：

$E_{z-P\left(z\right)}\[H\[p\left(y\right|D\left(z\right),D)\]\]\≈\frac{1}{M}\underset{i=1}{\overset{M}{\Σ}}H\[p\left(y\right|G\left(z^i\right),D\]---\left(3\right)$

Wherein, M represents the quantity of independent sample, noise vector zⁱ~P (z)；

2.4 calculate the entropy of classification y marginal probability in generation data set g according to formula (4)：

$H_g\[p\left(y\right|D)\]\≈H\[\frac{1}{M}\underset{i=1}{\overset{M}{\Σ}}p\left(y\right|G\left(z^i\right),D)\]---\left(4\right)$

2.5 unsupervised learnings are modeled, shown in the object function such as formula (5) of the non-supervision image classification based on confrontation generation network：

$\left\{\begin{array}{c}{L}_D=\underset{D}{\max }{H}_{data}\[p\left(y\right|D)\]-E_{x-data}\[H\[p\left(y\right|x,D)\]\]+E_{z~P\left(z\right)}\[H\[p\left(y\right|G\left(z\right),D)\]\]\\ L_G=\underset{G}{\min }-H_g\[p\left(y\right|D)\]+E_{z~P\left(z\right)}\[H\[P\left(y\right|G\left(z\right),D)\]\]\end{array}\right.---\left(5\right)$

Wherein, L_DIt is arbiter D loss function, L_GIt is maker G loss function；

2.6 semi-supervised learnings are modeled：When known to the classification information of part current image in training set data, above-mentioned unsupervised point Class problem can be further converted to semisupervised classification problem, now, and maker G loss function keeps constant in formula (5), i.e. L^label _G=L_G, and arbiter D loss function is adjusted to：

$\begin{array}{c}{L}_D^{label}=\underset{D}{\max }{H}_{data}\[p\left(y\right|D)\]-E_{x-data}\[H\[p\left(y\right|x,D)\]\]+E_{z~P\left(z\right)}\[H\[p\left(y\right|G\left(z\right),D)\]\]\\ +{\mathrm{\λE}}_{(x,y)~{\mathrm{data}}^{label}}\[CE\[y,p\left(y\right|x,D)\]\]\end{array}---\left(6\right)$

Wherein, λ is balance parameters, label data collection data^label={ (x¹,y¹),…,(x^L,y^L), L represents label data Number, CE [y, p (y | x, D)] represent data^labelThe condition distribution p (y | x, D) of middle image x true classification y and prediction classification Cross entropy, is specifically calculated as shown in formula (7)：

$CE\[y,p\left(y\right|x,D)\]=-\underset{i=1}{\overset{K}{\Σ}}{y}_i\log p(y=y_i|x,D)---\left(7\right)$

2.7 model optimizations are solved：Because maker G and arbiter D can be micro-, therefore can be according to object function (5) or (6) Using stochastic gradient descent method to L_DAnd L_GCarry out alternately training；

Step 3, measurement of rate of flow：The classification results of step 2 are corresponded with predefined flow velocity interval；

Step 4, state analysis：Abnormal state signal is sent when monitoring that water velocity exceeds predetermined threshold value.