CN106845528A - A kind of image classification algorithms based on K means Yu deep learning - Google Patents

Abstract

The invention discloses a kind of image classification algorithms based on K means Yu deep learning, including step：1) without label image as input picture, and image block composition size identical will be randomly selected without label image collection；2) Optimal cluster center is extracted using K means algorithms；3) construction feature mapping function, extracts the characteristics of image without label image collection；4) pondization operation and normalized are carried out；5) secondary Optimal cluster center is extracted using K means algorithms, and uses convolution operation, extract final image feature, final image feature is standardized；6) the final image feature by standardization is classified by sorter.The present invention has the advantages that simple, efficient, training parameter is few, has good effect for the classification of magnanimity dimensional images, and input picture is pre-processed, and reaches improvement image classification effect, improves the effect of nicety of grading.

Description

A kind of image classification algorithms based on K-means Yu deep learning

Technical field

Refer to a kind of image based on K-means Yu deep learning specifically the invention belongs to areas of information technology Sorting algorithm, it is adaptable to the classification of magnanimity high dimensional data image in internet, is additionally operable to network image retrieval, video frequency searching, distant The view data classification in the fields such as sense image classification, interactive entertainment, intelligent robot.

Background technology

In mass image data processing technology field, deep learning is a kind of relatively conventional algorithm.Deep learning conduct A kind of algorithm proposed in 2006 by Hinton, and have been widely recognized with application, its essence be by set up have it is many The artificial nerve network model of individual hidden layer and large-scale training data learn useful and abstract feature, final result It is the accuracy for lifting image classification.Therefore, deep learning can well solve mass image data process problem.Hinton is carried Go out DBN networks, and confirm that the shallower layer network structure of (1) deep layer network structure has more preferable feature learning ability；(2) pass through The mode successively trained enables that deep layer network structure is trained well.Hereafter, there is more deep learning model successive It is suggested, these models have also confirmed the viewpoint of Hinton.

The training method of traditional neural network is mainly by the way of back-propagation algorithm (BP), using random initializtion Method, the output of current network is calculated by way of iteration, then according between current predictive label and physical tags Difference goes constantly to adjust the parameter between preceding layers, until the convergence of whole model.Traditional BP algorithm has supervision to learn as one kind Practise algorithm, there are problems that gradient disperse problem, lack of training samples and, at the same time, due to internet Middle magnanimity is in explosive growth without label image, and traditional BP algorithm can not meet the need that magnanimity is classified without label image a few days ago Ask.

The content of the invention

The training method that the purpose of the present invention is directed to traditional neural network has gradient disperse, training sample not Be enough to and local optimum problem, propose a kind of image classification algorithms based on K-means Yu deep learning.

To achieve the above object, a kind of image classification algorithms based on K-means Yu deep learning designed by the present invention, Comprise the following steps：

1) without label image as input picture, and image block composition size identical will be randomly selected without label image Collection；

2) Optimal cluster center is extracted using K-means algorithms；

3) construction feature mapping function, extracts the characteristics of image without label image collection；

4) pondization operation and normalized are carried out；

5) secondary Optimal cluster center is extracted using K-means algorithms, and uses convolution operation, extract final image special Levy, final image feature is standardized；

6) the final image feature by standardization is classified by sorter.

Preferably, the step 2) specific steps include：

21) k initial cluster centre { μ of setting₁,μ₂,μ₃…μ_k, k is natural number, sets up the criterion function of initialization

Wherein, μ_jIt is each sample x⁽ⁱ⁾Corresponding cluster centre, j=1~k, i are natural number, and i ＞ j；x⁽ⁱ⁾Represent The n sample concentrated without label image, n is natural number, indicates the number that sample is concentrated without label image.

22) each sample x is asked for successively⁽ⁱ⁾To all initial cluster centre { μ₁,μ₂,μ₃…μ_kDistance minimum Value, is designated as the class label c of the sample⁽ⁱ⁾, by sample x⁽ⁱ⁾It is classified as c⁽ⁱ⁾Class, further according to the class label c⁽ⁱ⁾Update and calculate Cluster centre, obtains process cluster centre μ_j', j=1~k；

c⁽ⁱ⁾=argmin | | x⁽ⁱ⁾-μ_j′||

23) by all process cluster centre μ_j' bring criterion function calculating into, whether judgment criterion function restrains, no Then return to step 22), be then near step 24)；

24) by process cluster centre μ_j' it is defined as an Optimal cluster center { μ '₁,μ′₂,μ′₃…μ′_k, by each sample This x⁽ⁱ⁾Sort out to closest cluster centre, be designated as x_j ⁽ⁱ⁾, each sample x⁽ⁱ⁾To the class of closest cluster centre Distinguishing label is designated as c_j ⁽ⁱ⁾。

Preferably, the step 3) concretely comprise the following steps：Define sample x⁽ⁱ⁾Feature Mapping function, extract characteristic vector y⁽ⁱ⁾, wherein h (z) represents all samples in each class to the average value of the distance of cluster centre, Z_j ⁽ⁱ⁾Represent every in each class One sample x⁽ⁱ⁾To the distance of corresponding cluster centre；

y⁽ⁱ⁾=f_k(x)=max { 0, h (z)-Z_j ⁽ⁱ⁾}

Z_j ⁽ⁱ⁾=| | x⁽ⁱ⁾-μ_j||₂

Preferably, the step 4) in the formula of normalized be：

y⁽ⁱ⁾It is the characteristic vector of sample image block, var and mean represents variance and average, and σ is denoising constant,It is to return The characteristic vector of the image block after one change treatment.

Preferably, the step 5) detailed process include：

51) p initial cluster centre { μ of setting₁,μ₂,μ₃…μ_p, p is natural number, repeat step 2), obtain it is secondary most Good cluster centre { μ₁′,μ₂′,μ₃′…μ′_p}；

52) convolution operation is used, final image feature is extractedConvolution Formula is：

WhereinIt is final image feature, μ_lIt is secondary Optimal cluster center, l=1~p；

53) final image feature is normalized, the same step 4) of formula of normalized.

Preferably, the step 1) and step 2) between also include to being normalized and whitening processing without label image collection Pre-treatment step.

Preferably, the detailed process of the whitening processing includes：

A covariance matrix ∑) is calculated

Wherein,The sample concentrated without label image that indicates after normalized is represented, n is to indicate without label image Concentrate the number of sample；

B the characteristic vector U for) making covariance matrix ∑ is U=[u₁,u₂…u_n], U^TU=I, characteristic vector u₁,u₂…u_nStructure Into a base vector, for mapping data,Represent postrotational image, after input normalized without label image collection In sampleIt is expressed as by base vector of characteristic vector U

C the characteristic value for) setting covariance matrix ∑ is λ₂,…,λ_n, then the image after PCA albefactions be

Wherein ε is expressed as the constant of smoothed image block.

Advantages of the present invention includes：

(1) using unsupervised learning algorithm K-means as deep layer network structure training method, it is to avoid to each seed ginseng Several training, only needs training dictionary (i.e. cluster centre) so that training process is simple, and time efficiency is high, with it is simple, efficient, The advantages of training parameter is few, has good effect for the classification of magnanimity dimensional images.

(2) input picture is pre-processed, reaches improvement image classification effect, improve the effect of nicety of grading.

(3) normalization operation is taken, strengthens picture contrast to reduce the influence of light.

(4) due to there is certain correlation between view data, it is superfluous between image that the present invention takes PCA whitening processings to eliminate Yu Xing.

(5) average pondization is taken to process the dimension and integral image feature for reducing characteristic vector, after after the characteristics of image of pond, For the influence of each component of balance characteristics, the characteristics of image to pond is normalized, follow-up to improve Classifying quality of the Softmax graders to characteristics of image.

Brief description of the drawings

Fig. 1 is a kind of flow chart based on K-means Yu the image classification algorithms of deep learning of the present invention；

Fig. 2 is K-means algorithm flow charts；

Fig. 3 is that convolution extracts feature schematic diagram；

Fig. 4 a are the graph of a relation of the classification accuracy that three kinds of models are classified to MNIST data sets and iterations；

Fig. 4 b are the graph of a relation of the classification accuracy that three kinds of models are classified to Cifar-10 data sets and iterations；

Fig. 4 c are the pass of the classification accuracy that three kinds of models are classified to The four-vehicle data sets and iterations System's figure.

Specific embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is described in further detail.

As depicted in figs. 1 and 2, a kind of image classification algorithms based on K-means Yu deep learning of the present invention, including it is as follows Step：

1) without label image as input picture, and image block composition size identical will be randomly selected without label image Collection.It is 100,000 that the sample size without label image collection is set in the present embodiment, and corresponding size is 12*12*3 Image block.

To being pre-processed without label image collection, including normalized and whitening processing.

The process of normalized is：

Wherein, x is the sample without label image collection of input,Represent indicating without label image collection after normalized In sample, var and mean represents variance and average respectively, and σ is denoising constant, it is to avoid denominator is 0 and to image denoising.

The process of whitening processing is：

A covariance matrix ∑) is calculated

Wherein,The sample concentrated without label image that indicates after normalized is represented, n is to indicate without label image collection The number of middle sample, n=100 in this example, 000.

B the characteristic vector U for) making covariance matrix ∑ is U=[u₁,u₂…u_n], U^TU=I, characteristic vector u₁,u₂…u_nStructure Into a base vector, for mapping data,Represent postrotational image, after input normalized without label image collection In sampleIt is expressed as by base vector of characteristic vector U

C the characteristic value for) setting covariance matrix ∑ is λ₂,…,λ_n, then the image after PCA albefactions be

Wherein ε is expressed as the constant of smoothed image block.

2) Optimal cluster center is extracted using K-means algorithms.

Using pretreated without label image collection as K-means cluster datas, by K-means clustering algorithm training nets Network obtains cluster centre, that is, dictionary.K-means clustering algorithms are used as a kind of unsupervised learning algorithm, it is to avoid to various The training of parameter, only needs training dictionary so that training process is simple, and time efficiency is high.

It is { x without label image collection to make pretreated⁽¹⁾,x⁽²⁾,x⁽³⁾,…,x⁽ⁿ⁾},x⁽ⁱ⁾∈Rⁿ(wherein n represents image Concentrate the number of sample, i=1~n, n=100 in this example, 000, x⁽ⁱ⁾Represent some sample in n sample, RⁿRepresent n Dimensional vector), by K-means clustering algorithms to being clustered without label image collection.

21) k initial cluster centre { μ of setting₁,μ₂,μ₃…μ_k, k is natural number, and k=1600 is taken in this example, is set up The criterion function of initialization

Wherein, μ_jIt is each sample x⁽ⁱ⁾Corresponding cluster centre, j=1~k, i, j are natural number, and i ＞ j；x⁽ⁱ⁾Table Show the n sample concentrated without label image.

22) each sample x is asked for successively⁽ⁱ⁾To all initial cluster centre { μ₁,μ₂,μ₃…μ_kDistance minimum Value, is designated as the class label c of the sample⁽ⁱ⁾, by sample x⁽ⁱ⁾It is classified as c⁽ⁱ⁾Class, further according to the class label c⁽ⁱ⁾Update and calculate Cluster centre, obtains process cluster centre μ_j', j=1~k.

c⁽ⁱ⁾=argmin | | x⁽ⁱ⁾-μ_j′||

23) by all process cluster centre μ_j' bring criterion function calculating into, whether judgment criterion function is restrained, is otherwise returned Return step 22), be then near step 24).

24) by process cluster centre μ_j' it is defined as an Optimal cluster center { μ '₁,μ′₂,μ′₃…μ′_k, by each sample This x⁽ⁱ⁾Sort out to closest cluster centre, be designated as x_j ⁽ⁱ⁾, each sample x⁽ⁱ⁾To the class of closest cluster centre Distinguishing label is designated as c_j ⁽ⁱ⁾。

3) construction feature mapping function, extracts the characteristics of image without label image collection.

Define each sample x⁽ⁱ⁾Feature Mapping function, extract characteristic vector y⁽ⁱ⁾, wherein h (z) represented in each class All samples to the distance of cluster centre average value, Z_j ⁽ⁱ⁾Represent all samples in each class to corresponding cluster centre Distance；

y⁽ⁱ⁾=f_k(x)=max { 0, h (z)-Z_j ⁽ⁱ⁾}

Z_j ⁽ⁱ⁾=| | x⁽ⁱ⁾-μ_j||₂

When Feature Mapping function-output is 0, represent that this feature is more than " average value " to the distance of cluster centre.To every The image of one width input 64*64 sizes, the step-length s=1 of setting, selection receptive field size is (i.e. in the image block of 64*64 sizes The area size of selection) for 12*12 sample block as Feature Mapping function input, by sample block be mapped as k dimension feature Expression, the feature that a size is tieed up for the individual k of (64-12+1) * (64-12+1) can be obtained for every piece image.So ensure Most of characteristic values are exported for 0, makes it have openness, this rarefaction representation is widely applied in computer vision.

4) pondization operation and normalized are carried out.

After this, we are processed by pondization reduces the dimension of characteristic vector.

Because too high by image feature vector dimension that Feature Mapping function is obtained, it is unfavorable for Softmax graders pair Characteristics of image carries out classification and easily over-fitting occurs, so we take average pondization to process the characteristics of image for extracting reducing The dimension and integral image feature of characteristic vector.After after the characteristics of image of pond, for the influence of each component of balance characteristics, with Improve classifying quality of the follow-up Softmax graders to characteristics of image, our characteristics of image to pond are normalized place Reason.

The formula of normalized is：

It is the characteristic vector of the image block after standardization, y⁽ⁱ⁾The characteristic vector of sample image block, var and Mean represents variance and average, and σ is denoising constant.

5) secondary Optimal cluster center is extracted using K-means algorithms, and uses convolution operation, extract final image special Levy, final image feature is standardized.The image block characteristics of a*a sizes are obtained by training, we by its Convolution operation is carried out as convolution kernel and input picture, characteristics of image is extracted, as shown in Figure 2.

51) p initial cluster centre { μ of setting₁,μ₂,μ₃…μ_p, p is natural number, and p=2000 is taken in this example, is repeated Step 21)~23), by step 21)~23) in k replace with p, obtain secondary Optimal cluster center { μ₁′,μ₂′,μ₃′…μ ′_p}；

52) convolution operation is used, the final image feature of the image block of input is done into convolution with the cluster centre for extracting, carried Take final image featureConvolution Formula is：

WhereinIt is final image feature, μ_lIt is secondary Optimal cluster center, l=1~p；

53) final image feature is normalized, with the influence of each component of balance characteristics again, normalization The same step 4) of formula for the treatment of.

6) the final image feature by normalized is classified by Softmax sorters.

Experimental data：

A kind of image classification algorithms (KDL is expressed as in following table) based on K-means Yu deep learning proposed by the present invention, Compared with classical SAE, Stacked SAE algorithms.Experimental result is as shown in Table 1 to Table 3.

As can be seen from Table 1, to MNIST data sets classify in, set forth herein KDL category of model accuracys rate in iteration Maximum 97.52%, sparse self-encoding encoder model (SAE) and storehouse self-encoding encoder (Stacked SAE) point are reached at 300 times 4.52% and 16.8% is not higher by.From Fig. 4 a, KDL models are to MNIST data sets classification accuracy as iterations increases Plus SAE models and Stacked SAE models are far above all the time.Because SAE models can not effectively, accurately using single layer network structure Expression characteristics of image, so having the ability of preferably expression characteristics of image using K-means multitiered networks structures herein.It is real As shown by data KDL models entirety classification performance is tested better than SAE models, the experimental result it is anticipated that analyzing is also complied with.With Stacked SAE models are compared, KDL models have in training process and assorting process compared to Stacked SAE models it is simple, Efficiently, the advantages of training parameter is few, so its classification performance is better than Stacked SAE models.Experimental data equally confirms, KDL Model is far above Stacked SAE models in Classification and Identification accuracy rate.

Table 1. 3 kinds of models classification accuracy on MNIST data sets

As can be seen from Table 2, in classifying to Cifar-10 data sets, when iterations is 100, KDL categories of model are accurate True rate is 61.34%, and sparse self-encoding encoder model (SAE) and storehouse self-encoding encoder (Stacked SAE) are higher by respectively 1.65% and 31.02%.And knowable to from Fig. 4 b, as iterations increases, three kinds of category of model accuracys rate are in slow increasing Long status.Now, KDL Model Identifications rate is less than SAE models less than 1%, and is higher by Stacked SAE models 22.2%.Due to Identical experiment parameter is used when experimental result unavoidably has error and different pieces of information collection is classified, experiment may be caused to tie There is certain deviation in fruit.Integral experiment as shown by data, KDL models are higher than another two in the classification accuracy to Cifar-10 data sets Plant the classification accuracy of model.

Table 2. 3 kinds of models classification accuracy on Cifar-10 data sets

As can be seen from Table 3, in classifying to The four-vehicle data sets, KDL category of model recognition accuracies exist Maximum 80.87%, sparse self-encoding encoder model (SAE) and storehouse self-encoding encoder (Stacked are reached during iteration 100 times SAE 2.32% and 16.9%) is higher by respectively.Knowable to from Fig. 4 c, as iterations increases, the identification of KDL categories of model is accurate Rate is consistently higher than SAE models and Stacked SAE models.SAE categories of model accuracy rate is in as iterations increases and first increases Decline state after length, maximum 79.47% is reached when iterations is 400.Stacked SAE categories of model accuracys rate with Iterations increase is also at first to increase and declines state afterwards, and maximum 67.31% is reached when iterations is 200.In synthesis Stating analysis can obtain, and KDL categories of model recognition performance is better than SAE models and Stacked SAE models.

3. 3 kinds of models of table classification accuracy on The four-vehicle data sets

By three kinds of models, classification accuracy can be obtained on three data sets, set forth herein based on K-means and depth The KDL image classification models of study, are not only better than SAE models and Stacked SAE models in Classification and Identification accuracy rate, and Also have the advantages that K-means clustering algorithms it is simple, efficiently, learning parameter is few and deep learning has, and to be good at treatment big The ability of scale image.

Besides these examples, the present invention can also have other ways of realization, and all use equivalents or equivalent transformation are formed Scheme, all fall within the protection domain of this patent requirement.

The content not being described in detail in this specification belongs to prior art known to professional and technical personnel in the field.

Claims (7)

1. a kind of image classification algorithms based on K-means Yu deep learning, it is characterised in that：Comprise the following steps：

1) without label image as input picture, and image block composition size identical will be randomly selected without label image collection；

2) Optimal cluster center is extracted using K-means algorithms；

3) construction feature mapping function, extracts the characteristics of image without label image collection；

4) pondization operation and normalized are carried out；

5) secondary Optimal cluster center is extracted using K-means algorithms, and uses convolution operation, extract final image feature, it is right Final image feature is standardized；

6) the final image feature by standardization is classified by sorter.

2. a kind of image classification algorithms based on K-means Yu deep learning according to claim 1, it is characterised in that： 21) k initial cluster centre { μ of setting₁,μ₂,μ₃…μ_k, k is natural number, sets up the criterion function of initialization

$J\left(\mathrm{\μ}\right)=\underset{i=1}{\overset{n}{\Σ}}\left|\right|x^{\left(i\right)}-{\mathrm{\μ}}_j|{|}^2,$

Wherein, μ_jIt is each sample x⁽ⁱ⁾Corresponding cluster centre, j=1~k, i are natural number, and i ＞ j；x⁽ⁱ⁾Represent n nothing The sample that label image is concentrated, n is natural number, indicates the number that sample is concentrated without label image.

22) each sample x is asked for successively⁽ⁱ⁾To all initial cluster centre { μ₁,μ₂,μ₃…μ_kDistance minimum value, note It is the class label c of the sample⁽ⁱ⁾, by sample x⁽ⁱ⁾It is classified as c⁽ⁱ⁾Class, further according to the class label c⁽ⁱ⁾Update and calculate cluster Center, obtains process cluster centre μ_j', j=1~k；

c⁽ⁱ⁾=argmin | | x⁽ⁱ⁾-μ′_j||

${\mathrm{\μ}}_j=\frac{{\Σ}_{i=1}^{n}1\{c^{\left(i\right)}=j\}{x}^{\left(i\right)}}{{\Σ}_{i=1}^{n}1\{c^{\left(i\right)}=j\}}$

23) by all process cluster centre μ '_jBring criterion function calculating into, whether judgment criterion function is restrained, otherwise returned Step 22), be then near step 24)；

24) by process cluster centre μ '_jIt is defined as an Optimal cluster center { μ '₁,μ′₂,μ′₃…μ′_k, by each sample x⁽ⁱ⁾Sort out to closest cluster centre, be designated as x_j ⁽ⁱ⁾, each sample x⁽ⁱ⁾To the classification mark of closest cluster centre Label are designated as c_j ⁽ⁱ⁾。

3. a kind of image classification algorithms based on K-means Yu deep learning according to claim 2, it is characterised in that： The step 3) concretely comprise the following steps：Define sample x⁽ⁱ⁾Feature Mapping function, extract characteristic vector y⁽ⁱ⁾, wherein h (z) expressions In each class all samples to the distance of cluster centre average value, Z_j ⁽ⁱ⁾Represent each sample x in each class⁽ⁱ⁾To right The distance of the cluster centre answered；

$\begin{array}{c}{y}^{\left(i\right)}=f_k\left(x\right)=\max \{0,h\left(z\right)-Z_j^{\left(i\right)}\}\\ Z_j^{\left(i\right)}=\left|\right|x^{\left(i\right)}-{\mathrm{\μ}}_j|{|}_2\end{array}\mathrm{..}$

4. a kind of image classification algorithms based on K-means Yu deep learning according to claim 3, it is characterised in that： The step 4) in the formula of normalized be：

${\hat{y}}^{\left(i\right)}=(y^{\left(i\right)}-mean(y^{\left(i\right)}\left)\right)/\sqrt{\mathrm{var}\left(y^{\left(i\right)}\right)+\mathrm{\σ}}$

y⁽ⁱ⁾It is the characteristic vector of sample image block, var and mean represents variance and average, and σ is denoising constant,It is standardization The characteristic vector of the image block after treatment.

5. a kind of image classification algorithms based on K-means Yu deep learning according to claim 4, it is characterised in that： The step 5) detailed process include：

51) p initial cluster centre { μ of setting₁,μ₂,μ₃…μ_p, p is natural number, repeat step 2), obtain two suboptimums and gather Class center { μ '₁,μ′₂,μ′₃…μ′_p}；

52) convolution operation is used, final image feature is extractedConvolution Formula is：

${\tilde{y}}^{\left(i\right)}={\hat{y}}^{\left(i\right)}*{\mathrm{\μ}}_l$

WhereinIt is final image feature, μ_lIt is secondary Optimal cluster center, l=1~p；

53) final image feature is normalized, the same step 4) of formula of normalized.

6. a kind of image classification algorithms based on K-means Yu deep learning according to claim 1, it is characterised in that： The step 1) and step 2) between also include to be normalized without label image collection and whitening processing pre-treatment step.

7. a kind of image classification algorithms based on K-means Yu deep learning according to claim 6, it is characterised in that： The detailed process of the whitening processing includes：

A covariance matrix ∑) is calculated

Wherein,The sample concentrated without label image that indicates after normalized is represented, n is to indicate concentrate sample without label image This number,

B the characteristic vector U for) making covariance matrix ∑ is U=[u₁,u₂…u_n], U^TU=I, characteristic vector u₁,u₂…u_nConstitute one Base vector, for mapping data,Represent postrotational image, the sample concentrated without label image after input normalized ThisIt is shown as by base table of characteristic vector U

C the characteristic value for) setting covariance matrix ∑ is λ₂,…,λ_n, then the image after PCA albefactions be

Wherein ε is expressed as the constant of smoothed image block.