CN106096661A - Zero sample image sorting technique based on relative priority random forest - Google Patents

Abstract

The present invention proposes a kind of zero sample image sorting technique based on relative priority random forest, attribute sequence score model is set up according to the image that the relativeness between image category and image attributes is unknown classification, the score model that sorted by the attribute of all images trains random forest grader as training sample, and the label of test image is predicted by the random forest grader obtained finally according to the attribute sequence score and training testing image.The method of the present invention is capable of zero sample image classification, and has the advantages such as Classification and Identification rate is high, model stability is strong.

Description

Zero sample image sorting technique based on relative priority random forest

Technology zero territory

The invention belongs to pattern recognition zero territory, a kind of zero sample image classification based on relative priority random forest Method.

Background technology

1.1 0 sample image classification

Zero sample image classification is one of study hotspot of current area of pattern recognition, with traditional image classification problem not It is sorted in, with, zero sample image, the image that test phase classifies and identify and has neither part nor lot in the training of sorter model.Such as Fig. 1 institute Showing, the training stage, marked image covered " Lion ", and " Athletic shoes " and " Polar bear " three classifications are (the most Know classification), and the image of test phase occurs in that " Stiletto " classification (i.e. unknown classification), due to " Stiletto " class Bing Meiyou not participate in the training of grader, therefore grader will be unable to predict its label.Training data and the distribution of test data Difference makes zero sample image classification become extremely difficult learning tasks, but this problem scenes is widely present in calculating The fields such as the classification of machine vision, image, face and speech recognition.In zero sample image classification problem, in order to realize from known class Being clipped to the knowledge migration of unknown classification, disaggregated model is accomplished by building one from low-level image feature to classification mark by perceptual property The bridge signed.

1.2 relative priority study

Perceptual property (also referred to as attribute) refers to can be by artificial mark and the characteristic (example that can observe in the picture As, have wing, dark hair), attribute is broadly divided into two-value property and relative priority, wherein relative priority represent certain piece image with Other images compare the number containing a certain attribute.Owing to from the angle of cognition of the mankind, being often appreciated and understood by things It is to go to treat in the way of comparing, can more accurately express semantic attribute information hence with relative priority.Such as, from figure Can visually see in 2, (a) shows the attribute of significantly " young ", and (c) does not has " young " this attribute, for For (b), then can not describe with " young " or " not young " simply, and can be with the mode accurate description compared For: " (b) is more young than (c) and less young than (a) ".Therefore, relative priority can be more accurate Semantic information is expressed on ground, has stronger iamge description ability and interactive capability, it is possible to effectively downscaled images low-level image feature And the semantic gap between high-level semantics features.

Relative priority is in addition to the above-mentioned size that can be applied to description attribute intensity, additionally it is possible to by manually having Supervise and provide a feedback to grader^[And interactively choose the result of image retrieval, thus improve of Active Learning Habit ability.There is research to be mixed with relative priority by two-value property and propose one Spoken attributive classification device so that attribute can be with One more naturally mode describes image.Research is also had to be tied mutually with relative priority learning framework by degree of depth nerve convolutional network Close, to increase the precision of attribute sequence.In nearest research, relative priority is used for solving text and describes and zero sample learning Etc. problem, such as: first, Gaussian distribution model is set up for all of known class；Then, built by artificial selection's known class The Gaussian distribution model of vertical unknown classification；Finally, the method utilizing maximal possibility estimation carries out Tag Estimation to test image.So And, there is certain shortcoming in this method: (1) assumes that all known class images and the unknown equal Gaussian distributed of class image are not Reasonably；(2) owing to needing to select with manually having supervision classification in modeling process, therefore can artificially be led The impact of sight factor thus cause the accuracy of model the highest；(3) there is bigger error in maximum Likelihood, and this also will The accuracy of image classification is impacted.

Summary of the invention

Technical problem solved by the invention is to provide a kind of zero sample image based on relative priority random forest to divide Class method, sets up attribute sequence score mould according to the image that the relativeness between image category and image attributes is unknown classification Type, the score model that sorted by the attribute of all images trains random forest grader as training sample, finally according to test The label of test image is predicted by the random forest grader that the attribute sequence score of image and training obtain.This method It is capable of zero sample image classification, and Classification and Identification rate is high, model stability is strong.

The technical solution realizing the object of the invention is:

Zero sample image sorting technique based on relative priority random forest, comprises the following steps:

Step 1: the low-level image feature of given known class image and class label collection { x₁,x₂,...,x_S；y₁,y₂,...,y_S}、 Low-level image feature collection { the z of unknown class image₁,z₂,...,z_U, the orderly attribute of known class image is to collection { O₁,...,O_M, known class The like attribute of image is to collection { S₁,...,S_M, a number T of random tree and sampling percentage rate η, set up majorized function, wherein, S, U, M, T are positive integer, η ∈ (0,1)；

Step 2: utilize low-level image feature and the class label collection { x of known class image₁,x₂,...,x_S；y₁,y₂,...,y_S}、 Attribute is to collection { O in order₁,...,O_MAnd like attribute to collection { S₁,...,S_MSolving-optimizing function, obtain M attribute sequence letter NumberWherein w_mFor projection vector,For w_mTransposition, i=1,2 ..., s, m=1,2 ..., M；

Step 3: set up the attribute sequence score model of known class imageAttribute row with unknown class image Sequence score modelAnd form training sample set Ω, all images are positioned in attribute space, wherein,Represent known class image correspondence low-level image feature x respectively₁,x₂,...,x_SAttribute sequence score,Represent unknown class image correspondence low-level image feature z respectively₁,z₂,...,z_UAttribute sequence point；

Step 4: training sample set Ω is carried out T the Bootstrap stochastical sampling that percentage rate is η of sampling, is sampled Sample set Ω_t=BootstrapSampling (Ω), t=1,2 ..., T；

Step 5: generation random tree classification device:

Step 5-1: if Ω_tIn the classification of all samples identical, then present node is returned as leaf node, and according to sample This class label carrys out this node classification of labelling；Otherwise, step 5-2 is forwarded to；

Step 5-2: randomly choose parameter space subset:Γ(Ω_t) it is complete parameter space, Γ_sub (Ω_t) it is Γ (Ω_t) subset, for each parameter space subset Γ_sub(Ω_t), calculate information gain IG (θ_j|Ω_t), Optimized parameter to Weak Classifier:J=1,2 ..., | Γ_sub|, θ_jRepresent subset Γ_sub(Ω_t) In jth classification；

Step 5-3: the current data set making left and right child node is sky:

Step 5-4: according to optimized parameter θ^*Calculate Weak Classifier h (r_i|θ^*) value, if h (r_i|θ^*)=1, then by (r_i, y_i) add the data set of left child node: Ω to_left=Ω_left∪{(r_i,y_i)}；If h is (r_i|θ^*)=0, then by (r_i,y_i) add Data set to right child node: Ω_right=Ω_right∪{(r_i,y_i), wherein, r_iRepresent attribute sequence score, y_iRepresent classification Label；

Step 5-5: data set Ω_leftAnd Ω_rightBecome the child node of this node, these child nodes are repeated respectively to step Rapid 5-1 to 5-4, obtains the t random tree classification device；

Step 6: repetition step 4, to step 5, obtains zero sample image grader based on relative priority random forest TreeRoot₁,...,TreeRoot_T；

Step 7: utilize attribute ranking functionsCalculate attribute sequence score r of test image^(u)；

Step 8: by r^(u)Substitute into grader TreeRoot₁,...,TreeRoot_TIn, obtain r^(u)Belong to the probability of classification C, Calculate and export the class label of test image.

Further, the zero sample image sorting technique based on relative priority random forest of the present invention, excellent in step 1 Changing function is:

$f_{min}=\frac{1}{2}{||w_m||}_2^2+C(\underset{i,j}{\Σ}{\mathrm{\ξ}}_{ij}^2+\underset{i,j}{\Σ}{\mathrm{\γ}}_{ij}^2)$

$s.t.\left\{\begin{array}{cc}{w}_m^T{x}_i\≥w_m^T{x}_j+1-{\mathrm{\ξ}}_{ij}& \∀(i,j)\∈\{O_1,\mathrm{...},O_M\}\\ \left|w_m^T{x}_i-w_m^T{x}_j\right|\≤{\mathrm{\γ}}_{ij}& \∀(i,j)\∈\{S_1,\mathrm{...},S_M\}\\ {\mathrm{\ξ}}_{ij}\≥0& \\ {\mathrm{\γ}}_{ij}\≥0& \end{array}\right.$

Wherein, ξ_ijIt is ordered into attribute to { O₁,...,O_MNon-negative relaxation factor, γ_ijIt is that like attribute is to { S₁,..., S_MNon-negative relaxation factor, parameter C is used for weighing maximization Edge Distance and meets attribute to relativeness.

Further, the zero sample image sorting technique based on relative priority random forest of the present invention, step 2 solves Majorized function uses the edge 1/ that will sort | | w_m| | maximization, non-negative relaxation factor ξ_ijAnd γ_ijMinimize, thus obtain optimum throwing Shadow vector.

Further, the zero sample image sorting technique based on relative priority random forest of the present invention, step 3 is set up The attribute sequence score model of known class image comprises the following steps:

Step 3-1: by known class of all categories between relation object compare the relation between the image attributes belonging to such；

Step 3-2: calculate the attribute sequence score of all known class images r_i ^(s)The sequence score of every dimensional table diagram picture correspondence attribute；

Step 3-3: the attribute of all known class images sort to be grouped into attribute sequence score model

Further, the zero sample image sorting technique based on relative priority random forest of the present invention is unknown in step 3 Class image sets up relativeness by the relation between attribute and known class, sorts with the attribute that this sets up unknown class image Sub-model, specifically includes following three kinds of situations:

(1) if there is known classM-th attributeWith unknown classM-th attributeSimultaneously Meet:And known classWithIt is and unknown classTwo known class that relative priority sequence is nearest, then scheme As the m-th attribute of model sorts to be divided into:

$r_{m,j}^{\left(u\right)}=\frac{1}{2}(r_{m,i}^{\left(s\right)}+r_{m,k}^{\left(s\right)})$

Wherein, i=1,2 ..., I, k=1,2 ..., K, j=1,2 ..., I K, I and K are known class respectivelyWithTotal number of images,It is respectively known classM-th attribute sequence score；

(2) if unknown classIt is in border and there is known classM-th attribute meetThen iconic model M-th attribute sort to be divided into:

$r_{m,j}^{\left(u\right)}=r_{m,i}^{\left(s\right)}-d_{m,i}^{\left(s\right)}$

Wherein, i=1,2 ..., I, j=1,2 ..., I,Represent between the attribute sequence score of training class image Mean difference, Represent the average of m-th attribute,

(3) if unknown classIt is in border and there is known classM-th attribute meetThen iconic model M-th attribute sort to be divided into:

$r_{m,j}^{\left(u\right)}=r_{m,k}^{\left(s\right)}+d_{m,k}^{\left(s\right)}$

Wherein, k=1,2 ..., K, j=1,2 ..., K,

Further, the zero sample image sorting technique based on relative priority random forest of the present invention, r in step 8^(u) The probability belonging to classification c is:

$p\left(c\right|r^{\left(u\right)})=\frac{1}{T}\underset{t=1}{\overset{T}{\Σ}}{p}_t\left(c\right|r^{\left(u\right)})$

Wherein, T is a number of random tree, p in forest_t(c|r^(u)) it is the categorical distribution of leaf node.

Further, the zero sample image sorting technique based on relative priority random forest of the present invention, step 8 is tested The class label of image is:

$\hat{c}=\underset{c\∈\{1,\mathrm{...},U\}}{\arg \max }p\left(c\right|r^{\left(u\right)}).$

The present invention uses above technical scheme compared with prior art, has following technical effect that

1, the zero sample image sorting technique based on relative priority random forest of the present invention is that each image is individually built Vertical attribute sequence score model so that the model participating in classifier training is the most reasonable and accurate；

2, the zero sample image sorting technique based on relative priority random forest of the present invention automatically selects known class Model is set up, it is to avoid the subjective impact that artificial selection's known class is brought for unknown classification；

3, the zero sample image sorting technique based on relative priority random forest of the present invention uses random forest grader Reduce error in classification, thus improve the accuracy of zero sample image classification；

4, the zero sample image sorting technique Classification and Identification rate based on relative priority random forest of the present invention is high, model is steady Qualitative by force.

Accompanying drawing explanation

Fig. 1 is zero sample image classification schematic diagram；

Fig. 2 is relative priority schematic diagram；

Fig. 3 is zero sample image sorting technique structured flowchart based on relative priority random forest；

Fig. 4 is the flow chart of study attribute ranking functions；

Fig. 5 be ordered into attribute to and like attribute to schematic diagram；

Fig. 6 is attribute ranking functions schematic diagram；

Fig. 7 is the flow chart setting up AR model；

Fig. 8 is the flow chart of training random forest grader；

Fig. 9 is the flow chart of prediction test image category label.

Detailed description of the invention

Embodiments of the present invention are described below in detail, and the example of described embodiment is shown in the drawings, the most ad initio Represent same or similar element to same or similar label eventually or there is the element of same or like function.Below by ginseng The embodiment examining accompanying drawing description is exemplary, is only used for explaining the present invention, and is not construed as limiting the claims.

Zero sample image sorting technique based on relative priority random forest, as it is shown on figure 3, comprise the following steps:

Step 1: such as (1) in Fig. 4, the low-level image feature of given known class image and class label collection { x₁,x₂,...,x_S；y₁, y₂,...,y_S, the low-level image feature collection { z of unknown class image₁,z₂,...,z_U, the orderly attribute of known class image is to collection {O₁,...,O_M, the like attribute of known class image is to collection { S₁,...,S_M, a number T of random tree and sampling percentage rate η, its In, S, U, M, T are positive integer, η ∈ (0,1), set up following majorized function:

$f_{min}=\frac{1}{2}{||w_m||}_2^2+C(\underset{i,j}{\Σ}{\mathrm{\ξ}}_{ij}^2+\underset{i,j}{\Σ}{\mathrm{\γ}}_{ij}^2)$

$s.t.\left\{\begin{array}{cc}{w}_m^T{x}_i\≥w_m^T{x}_j+1-{\mathrm{\ξ}}_{ij}& \∀(i,j)\∈\{O_1,\mathrm{...},O_M\}\\ \left|w_m^T{x}_i-w_m^T{x}_j\right|\≤{\mathrm{\γ}}_{ij}& \∀(i,j)\∈\{S_1,\mathrm{...},S_M\}\\ {\mathrm{\ξ}}_{ij}\≥0& \\ {\mathrm{\γ}}_{ij}\≥0& \end{array}\right.$

Wherein, ξ_ijIt is ordered into attribute to { O₁,...,O_MNon-negative relaxation factor, γ_ijIt is that like attribute is to { S₁,..., S_MNon-negative relaxation factor, parameter C is used for weighing maximization Edge Distance and meets attribute to relativeness.

Concrete principle is as follows:

Given training image collection I={i}, each image characteristic vector x_i∈R^dRepresent；The given genus with M attribute Property collectionFor each attribute a_m, given a series of orderly attributes are to O_m={ (i, j) } and like attribute are to S_m= (i, j) }, wherein,Represent that image i contains attribute more than image j；Represent image i Similar to image j containing attribute.Fig. 5 be orderly attribute as a example by " laughing at " this attribute to and like attribute to schematic diagram, its In the image attributes intensity of orderly attribute centering in different size, there is certain strong or weak relation, and the figure of like attribute centering As attribute intensity difference is few, there is not attribute strong or weak relation.The purpose of study attribute ranking functions is to obtain M attribute row Order function:

$r_m\left(x_i\right)=w_m^T{x}_i$

For m=1 ..., M, meet following restriction most possibly:

$\∀(i,j)\∈O_m:w_m^T{x}_i>w_m^T{x}_j$

$\∀(i,j)\∈S_m:w_m^T{x}_i=w_m^T{x}_j$

Wherein, w_mBeing projection vector, therefore, study attribute ranking functions is intended in low-level image feature space find the throwing of optimum Shadow direction so that the projection in the direction of all images has correct sequence.For solving the problems referred to above, introduce non-negative and relax Variable ξ_ijAnd γ_ij, obtain following majorized function:

$f_{min}=\frac{1}{2}{||w_m||}_2^2+C(\underset{i,j}{\Σ}{\mathrm{\ξ}}_{ij}^2+\underset{i,j}{\Σ}{\mathrm{\γ}}_{ij}^2)$

$s.t.\left\{\begin{array}{cc}{w}_m^T{x}_i\≥w_m^T{x}_j+1-{\mathrm{\ξ}}_{ij}& \∀(i,j)\∈\{O_1,\mathrm{...},O_M\}\\ \left|w_m^T{x}_i-w_m^T{x}_j\right|\≤{\mathrm{\γ}}_{ij}& \∀(i,j)\∈\{S_1,\mathrm{...},S_M\}\\ {\mathrm{\ξ}}_{ij}\≥0& \\ {\mathrm{\γ}}_{ij}\≥0& \end{array}\right.$

Wherein, ξ_ijIt is ordered into attribute to O_mThe non-negative relaxation factor of={ (i, j) }, γ_ijIt is that like attribute is to S_m=(i, J) non-negative relaxation factor }, parameter C is used for weighing maximization Edge Distance and meeting attribute to relativeness.

Step 2: utilize low-level image feature and the class label collection { x of known class image₁,x₂,...,x_S；y₁,y₂,...,y_S}、 Attribute is to collection { O in order₁,...,O_MAnd like attribute to collection { S₁,...,S_MSolving-optimizing function, will sort edge 1/ | | w_m|| Maximization, non-negative relaxation factor ξ_ijAnd γ_ijMinimize, obtain optimum projection vector, thus obtain M attribute ranking functionsWherein w_mFor projection vector,For w_mTransposition, i=1,2 ..., s, m=1,2 ..., M.Set up attribute row Order function is actually the function that training image can be sorted by study one exactly, such as Fig. 4 (2), the edge limit of sequence System allows the distance between two images nearest in whole sequence maximum exactly.As shown in Figure 6, the purpose of sequence is by data point (representing with 1,2,3,4,5,6 respectively) preferably sorts, and allows in queue the edge between two nearest data points (2,3) Greatly, therefore ranking functions can preferably represent the relativeness of attribute intensity.

Step 3: set up the attribute sequence score model of known class imageAttribute row with unknown class image Sequence score modelAnd form training sample set Ω, all images are positioned in attribute space, wherein,Represent known class image correspondence low-level image feature x respectively₁,x₂,...,x_SAttribute sequence score,Represent unknown class image correspondence low-level image feature z respectively₁,z₂,...,z_UAttribute sequence point, such as Fig. 7 (3) institute Show.

Assuming that having S class image is known class, having U class image is unknown classification, general by zero sample image classification Reading and understand, the image of S class known class can directly participate in the training of grader, and the image of U class the unknown classification can not be direct Participate in the training of grader, can only appear on the test phase of zero sample image classification, therefore sort at the attribute setting up image During score model, it is known that the method that the image of the image of classification and unknown classification is used is different, and the present invention proposes to adopt Set up using the following method image attribute sequence score model:

Initially set up the attribute sequence score model of known class image, comprise the following steps, such as Fig. 7 (1):

Step 3-1: by known class of all categories between relation object compare the relation between the image attributes belonging to such, That is: for arbitrary attribute a_mAnd known classWith

Step 3-2: calculate the attribute sequence score of all known class images The sequence score of every dimensional table diagram picture correspondence attribute；

Step 3-3: the attribute of all known class images sort to be grouped into attribute sequence score model

The known class image x thus script d dimensional feature vector represented_iBy attribute sequence score r of M dimension_iRepresent: x_i∈R^d→r_i∈R^M, wherein,r_iThe sequence of every dimensional table diagram picture correspondence attribute Score.

Owing to unknown classification can not directly participate in the training process of grader, therefore can not be by same method to unknown class Image be modeled, but, unknown class image sets up relativeness by the relation between attribute and known class, such as " Bears " (unknown classification) is more than " giraffe " (known class) hair, but be not as many as " rabbit " (known class) hair.Such as Fig. 7 (2), specifically come Say, for attribute a_m, unknown classKnown class can be usedWithPoint following three kinds of situations carry out associated description:

(1) if there is known classM-th attributeWith unknown classM-th attributeSimultaneously Meet:And known classWithIt is and unknown classTwo known class that relative priority sequence is nearest, then scheme As the m-th attribute of model sorts to be divided into:

$r_{m,j}^{\left(u\right)}=\frac{1}{2}(r_{m,i}^{\left(s\right)}+r_{m,k}^{\left(s\right)})$

Wherein, i=1,2 ..., I, k=1,2 ..., K, j=1,2 ..., I K, I and K are known class respectivelyWithTotal number of images,It is respectively known classM-th attribute sequence score；

(2) if unknown classIt is in border and there is known classM-th attribute meetThen iconic model M-th attribute sort to be divided into:

$r_{m,j}^{\left(u\right)}=r_{m,i}^{\left(s\right)}-d_{m,i}^{\left(s\right)}$

Wherein, i=1,2 ..., I, j=1,2 ..., I,Represent between the attribute sequence score of training class image Mean difference, Represent the average of m-th attribute,

(3) if unknown classIt is in border and there is known classM-th attribute meetThen iconic model M-th attribute sort to be divided into:

$r_{m,j}^{\left(u\right)}=r_{m,k}^{\left(s\right)}+d_{m,k}^{\left(s\right)}$

Wherein, k=1,2 ..., K, j=1,2 ..., K,

The present invention uses following strategy to automatically select suitable known class and sets up the attribute sequence score mould of unknown class Type: prioritizing selection meetsKnown classWithAndWithIt is and unknown classRelative priority sorts Two near classes；IfIt is in border, does not meetKnown classWithThen select relative priority sequence The highestOr it is minimumTo unknown class modeling.

Step 4: training sample set Ω is carried out T the Bootstrap stochastical sampling that percentage rate is η of sampling, is sampled Sample set Ω_t=BootstrapSampling (Ω), t=1,2 ..., T.

Step 5: generate random tree classification device, each joint as shown in Figure 8, in random forest grader, in each tree Point can be regarded as a Weak Classifier, and (the training sample set Ω arriving this node is included known class sample setWith unknown class sample set) it is calculated a sorting criterion h (r | θ) ={ 0,1}, r ∈ R^MRepresenting a training sample, θ={ φ, ψ } is the parameter of this Weak Classifier, and wherein φ () is screening Function, ψ is a parameter matrix.

Step 5-1: if Ω_tIn the classification of all samples identical, then present node is returned as leaf node, and according to sample This class label carrys out this node classification of labelling；Otherwise, step 5-2 is forwarded to；

Step 5-2: randomly choose parameter space subset:Γ(Ω_t) it is complete parameter space, Γ_sub (Ω_t) it is Γ (Ω_t) subset, for each node Γ_subAll randomly choosing from Γ, this embodies in node split mistake Randomness in journey, to each parameter space subset Γ_sub(Ω_t), calculate information gain IG (θ_j|Ω_t), information gain is weighed The fall of training sample impurity level after division, can be defined as:

$IG\left(\mathrm{\θ}\right|\mathrm{\Ω})=H\left(\mathrm{\Ω}\right)-\underset{i\∈\{left,right\}}{\Σ}\frac{\left|{\mathrm{\Ω}}_i\left(\mathrm{\θ}\right)\right|}{\left|\mathrm{\Ω}\right|}H\left({\mathrm{\Ω}}_i\right(\mathrm{\θ}\left)\right)$

$H\left(\mathrm{\Ω}\right)=-\underset{c=1}{\overset{N_c}{\Σ}}p\left(c\right|\mathrm{\Ω})logp\left(c\right|\mathrm{\Ω})$

Wherein, N_cThe classification number of expression training sample, and p (c | Ω) represent the ratio shared by classification c in training sample set Ω Example,Represent the set of all training samples falling into this node, y_iRepresent the label of i-th sample, Ω_left (θ) and Ω_right(θ) it is illustrated respectively in parameter θ and falls into the sample set of left and right child node,Represent the unit in set omega Element number, H (Ω) represents the impurity level of the sample set falling into a node, describes by comentropy.

Then the optimized parameter of Weak Classifier is obtained:J=1,2 ..., | Γ_sub|, θ_jTable Show subset Γ_sub(Ω_tJth classification in).

It follows that " optimum " parameter θ of each node^*Node impurity level fall after cleaving should be made maximum.

Step 5-3: the current data set making left and right child node is sky:

Step 5-4: according to optimized parameter θ^*Calculate Weak Classifier h (r_i|θ^*) value, if h (r_i|θ^*)=1, then by (r_i, y_i) add the data set of left child node: Ω to_left=Ω_left∪{(r_i,y_i)}；If h is (r_i|θ^*)=0, then by (r_i,y_i) add Data set to right child node: Ω_right=Ω_right∪{(r_i,y_i), wherein, r_iRepresent attribute sequence score, y_iRepresent classification Label；

Step 5-5: data set Ω_leftAnd Ω_rightBecome the child node of this node, these child nodes are repeated respectively to step Rapid 5-1 to 5-5, obtains the t random tree classification device, it may be assumed that at each leaf node, is concentrated by statistics training sample and arrives this The rectangular histogram of the tag along sort of leaf node, can estimate the class distribution on this leaf node.Such repetitive exercise process is held always Row is to can not obtaining bigger information gain by continuing division.

Step 6: repetition step 4, to step 5, obtains zero sample image grader based on relative priority random forest TreeRoot₁,...,TreeRoot_T。

Step 7: such as Fig. 9, utilize attribute ranking functionsCalculate attribute sequence score r of test image^(u), The most all test images are all the images of unknown classification, are not engaged in the training of random forest grader..

Step 8: such as Fig. 9, by r^(u)Substitute into grader TreeRoot₁,...,TreeRoot_TIn, random at each iteratively Tree is carried out or the branch of left or right, until arrive each random tree leaf node, on each leaf node classification distribution namely It it is the classification results made of this tree.Classification in each leaf nodes is distributed and is averaged, i.e. can get r^(u)Belong to classification The probability of c:Wherein, T is a number of random tree, p in forest_t(c|r^(u)) it is the class of leaf node It is not distributed.Then calculate and export test image class label:

The above is only the some embodiments of the present invention, it is noted that for the ordinary skill people in this technology zero territory For Yuan, under the premise without departing from the principles of the invention, it is also possible to make some improvement, these improve the guarantor that should be regarded as the present invention Protect scope.

Claims (7)

1. zero sample image sorting technique based on relative priority random forest, it is characterised in that comprise the following steps:

Step 1: the low-level image feature of given known class image and class label collection { x₁, x₂..., x_S；y₁, y₂..., y_S, unknown Low-level image feature collection { the z of class image₁, z₂..., z_U, the orderly attribute of known class image is to collection { O₁..., O_M, known class image Like attribute to collection { S₁..., S_M, a number T of random tree and sampling percentage rate η, set up majorized function, wherein, S, U, M, T It is positive integer, η ∈ (0,1)；

Step 2: utilize low-level image feature and the class label collection { x of known class image₁, x₂..., x_S；y₁, y₂..., y_S, in order Attribute is to collection { O₁..., O_MAnd like attribute to collection { S₁..., S_MSolving-optimizing function, obtain M attribute ranking functionsWherein w_mFor projection vector,For w_mTransposition, i=1,2 ..., s, m=1,2 ..., M；

Step 3: set up the attribute sequence score model of known class imageSort with the attribute of unknown class image Sub-modelAnd form training sample set Ω, all images are positioned in attribute space, wherein,Represent known class image correspondence low-level image feature x respectively₁, x₂..., x_SAttribute sequence score,Represent unknown class image correspondence low-level image feature z respectively₁, z₂..., z_UAttribute sequence point；

Step 4: training sample set Ω is carried out T the Bootstrap stochastical sampling that percentage rate is η of sampling, obtains sample Collection Ω_t=BootstrapSampling (Ω), t=1,2 ..., T；

Step 5: generation random tree classification device:

Step 5-1: if Ω_tIn the classification of all samples identical, then present node is returned as leaf node, and according to sample Class label carrys out this node classification of labelling；Otherwise, step 5-2 is forwarded to；

Step 5-2: randomly choose parameter space subset:Γ(Ω_t) it is complete parameter space, Γ_sub(Ω_t) For Γ (Ω_t) subset, for each parameter space subset Γ_sub(Ω_t), calculate information gain/G (θ_j|Ω_t), obtain weak The optimized parameter of grader:J=1,2 ..., | Γ_sub|, θ_jRepresent subset Γ_sub(Ω_tIn) Jth classification；

Step 5-3: the current data set making left and right child node is sky:

Step 5-4: according to optimized parameter θ^*Calculate Weak Classifier h (r_i|θ^*) value, if h (r_i|θ^*)=1, then by (r_i, y_i) add Data set to left child node: Ω_left=Ω_left∪{(r_i, y_i)}；If h is (r_i|θ^*)=0, then by (r_i, y_i) add right sub-joint to The data set of point: Ω_right=Ω_right∪{(r_i, y_i), wherein, r_iRepresent attribute sequence score, y_iRepresent class label；

Step 5-5: data set Ω_leftAnd Ω_rightBecome the child node of this node, 5-is respectively repeated steps for these child nodes 1 to 5-4, obtains the t random tree classification device；

Step 6: repetition step 4, to step 5, obtains zero sample image grader based on relative priority random forest TreeRoot₁..., TreeRoot_T；

Step 7: utilize attribute ranking functionsCalculate attribute sequence score r of test image^(u)；

Step 8: by r^(u)Substitute into grader TreeRoot₁..., TreeRoot_TIn, obtain r^(u)Belong to the probability of classification C, calculate And export the class label of test image.

Zero sample image sorting technique based on relative priority random forest the most according to claim 1, it is characterised in that Majorized function in step 1 is:

$f_{min}=\frac{1}{2}{||w_m||}_2^2+C(\underset{i,j}{\Σ}{\mathrm{\ξ}}_{ij}^2+\underset{i,j}{\Σ}{\mathrm{\γ}}_{ij}^2)$

$s.t.\left\{\begin{array}{cc}{w}_m^T{x}_i\≥w_m^T{x}_j+1-{\mathrm{\ξ}}_{ij}& \∀(i,j)\∈\{O_1,\mathrm{...},O_M\}\\ \left|w_m^T{x}_i-w_m^T{x}_j\right|\≤{\mathrm{\γ}}_{ij}& \∀(i,j)\∈\{S_1,\mathrm{...},S_M\}\\ {\mathrm{\ξ}}_{ij}\≥0& \\ {\mathrm{\γ}}_{ij}\≥0& \end{array}\right.$

Wherein, ξ_ijIt is ordered into attribute to { O₁..., O_MNon-negative relaxation factor, γ_ijIt is that like attribute is to { S₁..., S_M? Non-negative relaxation factor, parameter C is used for weighing maximization Edge Distance and meeting attribute to relativeness.

Zero sample image sorting technique based on relative priority random forest the most according to claim 1, it is characterised in that In step 2, solving-optimizing function uses the edge 1/ that will sort | | w_m| | maximization, non-negative relaxation factor ξ_ijAnd γ_ijMinimize, from And obtain optimum projection vector.

Zero sample image sorting technique based on relative priority random forest the most according to claim 1, it is characterised in that The attribute sequence score model setting up known class image in step 3 comprises the following steps:

Step 3-1: by known class of all categories between relation object compare the relation between the image attributes belonging to such；

Step 3-2: calculate the attribute sequence score of all known class images 's The sequence score of every dimensional table diagram picture correspondence attribute；

Step 3-3: the attribute of all known class images sort to be grouped into attribute sequence score model

Zero sample image sorting technique based on relative priority random forest the most according to claim 1, it is characterised in that In step 3, unknown class image sets up relativeness by the relation between attribute and known class, sets up unknown class image with this Attribute sequence score model, specifically include following three kinds of situations:

(1) if there is known classM-th attributeWith unknown classM-th attributeMeet simultaneously:And known classWithIt is and unknown classTwo known class that relative priority sequence is nearest, then image mould The m-th attribute of type sorts to be divided into:

$r_{m,j}^{\left(u\right)}=\frac{1}{2}(r_{m,i}^{\left(s\right)}+r_{m,k}^{\left(s\right)})$

Wherein, i=1,2 ..., I, k=1,2 ..., K, j=1,2 ..., I K, I and K are known class respectivelyWithFigure As sum,It is respectively known classM-th attribute sequence score；

(2) if unknown classIt is in border and there is known classM-th attribute meetThen the of iconic model M attribute sorts to be divided into:

$r_{m,j}^{\left(u\right)}=r_{m,i}^{\left(s\right)}-d_{m,i}^{\left(s\right)}$

Wherein, i=1,2 ..., I, j=1,2 ..., I,Represent training class image attribute sequence score between average Difference, Represent the average of m-th attribute,

(3) if unknown classIt is in border and there is known classM-th attribute meetThen the of iconic model M attribute sorts to be divided into:

$r_{m,j}^{\left(u\right)}=r_{m,k}^{\left(s\right)}+d_{m,k}^{\left(s\right)}$

Wherein, k=1,2 ..., K, j=1,2 ..., K,

Zero sample image sorting technique based on relative priority random forest the most according to claim 1, it is characterised in that R in step 8^(u)The probability belonging to classification c is:

$p\left(c\right|r^{\left(u\right)})=\frac{1}{T}\underset{t=1}{\overset{T}{\Σ}}{p}_t\left(c\right|r^{\left(u\right)})$

Wherein, T is a number of random tree, p in forest_t(c|r^(u)) it is the categorical distribution of leaf node.

Zero sample image sorting technique based on relative priority random forest the most according to claim 6, it is characterised in that The class label testing image in step 8 is:

$\hat{c}=\underset{c\∈\{1,\mathrm{...},U\}}{\arg \max }p\left(c\right|r^{\left(u\right)}).$