CN110472657B - Image classification method based on trust function theory - Google Patents

Abstract

The invention discloses an image classification method based on trust function theory, which divides all images in an image set X to be classified into a plurality of initial subclasses; calculates the density value of each initial subclass; The size relationship of the initial single-class density value, the initial composite class and the images in the corresponding initial single-class are merged to generate a new single-class and a new composite class, until the sum of the number of new single-class and initial single-class generated after merging is c; The images in the initial composite class are divided into new single class, initial single class or new composite class to obtain the classification result of the images in the image set X to be classified; the present invention uses the density of different classes to divide the composite class and the single class into one The given rules are gradually merged together until the initial number of single classes after merging is exactly equal to the actual number of image classifications, which effectively reduces the risk of wrongly dividing data and avoids making wrong or biased decisions.

Description

Image Classification Method Based on Trust Function Theory

【Technical field】

The invention belongs to the technical field of image classification and recognition, and in particular relates to an image classification method based on trust function theory.

【Background technique】

Imbalanced data clustering is one of the important branches of pattern recognition and unsupervised machine learning, and has a wide range of applications in various fields.

Image recognition of aerial targets is an important research direction of aviation military surveillance system. Due to man-made confrontation, complex and changeable battlefield environment, and sensor performance limitations, when the image sensor captures images of moving targets in the air, there will be an imbalance in the number of images of different types of moving targets. At the same time, the images of different targets under certain viewing angles may be very similar, which makes the acquisition of target images with great uncertainty.

When the existing classical clustering methods perform unsupervised classification (clustering) on these unbalanced samples, it is often difficult to obtain reliable results, and even a "balance effect" occurs, that is, the number of images of different targets is very different. There may be very few images of one kind of target, and the other is very large, but after clustering, the image set will still be clustered into two target categories with roughly equal numbers, resulting in the occurrence of image classification. higher error rate. At the same time, the traditional classical clustering method does not consider that the images of different targets may be very similar in certain perspectives, that is, the images of different targets may be similar in some positions. When the similarity of images belonging to different targets is high , it is often difficult to accurately cluster these overlapping images.

[Content of the invention]

The purpose of the present invention is to provide an image classification method based on the trust function theory. When classifying images of different targets, the images are classified by using the densities of various types of images as classification criteria to improve the accuracy of image classification.

The present invention adopts the following technical solutions: an image classification method based on the trust function theory, comprising the following steps:

Divide all images in the image set X to be classified into multiple initial subclasses; wherein, the initial subclass consists of N initial single classes and multiple initial composite classes, and each initial composite class consists of two initial single classes, N≥c, c represents the number of true categories of all images in the image set X;

Calculate the density value of each initial subclass; where the density is the reciprocal of the average Euclidean distance of all images in the initial subclass;

According to the size relationship between the initial composite class and the density value of the corresponding initial single class, the images in the initial composite class and the corresponding initial single class are merged to generate a new single class and a new composite class, until the new single class and the initial single class are generated after merging The sum of the quantities is c;

The images in the initial composite class are divided into new single class, initial single class or new composite class, and the classification result of the images in the image set X to be classified is obtained.

Further, the magnitude relationship between the initial composite class and the corresponding initial single class density value consists of the following three cases:

Case 1: ρ _k,t ≥ρ _k and ρ _k,t ≥ρ _t ;

Case 2: ρ _k ≤ρ _k,t ≤ρ _t ;

Case 3: ρ _k,t ≤ρ _k and ρ _k,t ≤ρ _t ;

Among them, ρ _k is the density value of the k-th initial single class, 0≤k≤N, ρ _t is the density value of the t-th initial single class, 0≤t≤N, ρ _k,t is the k-th initial single class The density value of the initial composite class consisting of a single class and the t-th initial single class.

Further, the order of combining the images in the initial composite class and the corresponding initial single class is as follows: Case 1, Case 2, Case 3.

Further, the specific method for merging the images in the initial composite class and the corresponding initial single class is as follows:

Calculate the density distance value between the two initial single classes corresponding to the initial composite class, sort the initial composite classes according to the density distance value from small to large, and merge the images in the initial composite class and the corresponding initial single class into a new one. single class;

When the density distance values between the two initial single classes corresponding to the initial composite class are equal, the images in the initial composite class and the corresponding initial single class are merged into one according to the order of the density value of each initial composite class from large to small. new single class;

Among them, the density distance value is the difference between the ordinal numbers of the two initial single classes in the initial single class density set D _s , the initial single class density set D _s is the set of density values of N initial single class, and the initial single class density The set D _s is sorted according to the density value of the initial single class from large to small.

Further, when a new single class is generated and before the next initial composite class merging, it is judged whether the new single class and other new single classes have the same image. All images in the new single class are merged to generate a new single class.

Further, the specific method for dividing the images in the initial composite class into the new single class, the initial single class or the new composite class is as follows:

Select the image _xi in the initial composite class, and find the K ₂ images with the closest Euclidean distance in the image set X to be classified as its neighbor images, and generate K ₂ vectors, according to the image _xi and its neighbor images The Euclidean distance between generates K weights of ₂ vectors;

Calculate the sum vector of K ₂ vectors as the first sum vector;

Calculate the sum vector of the vectors of the vector end points corresponding to the vector end points in the K ₂ vectors that belong to the same initial single class/new single class, as the second sum vector;

Calculate the cosine value of the angle formed by each second sum vector and the first sum vector, and divide the image x _i into the initial single class/new single class corresponding to the second sum vector with the smallest cosine value of the included angle;

Repeat the above steps until all images in the initial composite class are divided.

Further, when the difference between the cosine value of the angle formed by the two second sum vectors and the first sum vector is smaller than the threshold, the image is divided into the initial single class/new single class formed by the two second sum vectors corresponding to in the new composite class.

Further, all images in the image set X to be classified are divided into a plurality of initial sub-classes, and the c-means clustering algorithm is specifically adopted.

Further, the image set X to be classified is an image set of asymmetric data.

Further, the density of each initial subclass is calculated using the formula:

表示初始子类A_j中的图像x_i与其在图像集X中的K₁个近邻的欧氏距离的平均距离，K₁为常数，n_j为初始子类A_j中图像的数量，i为初始子类A_j中图像的序数。where ρ _j is the density of the j-th initial subclass A _j ,

represents the average distance of the Euclidean distance between the image x _i in the initial subclass A _j and its K ₁ nearest neighbors in the image set X, where K ₁ is a constant, n _j is the number of images in the initial subclass A _j , and i is The ordinal number of images in the initial subclass A _j .

The beneficial effects of the present invention are as follows: the present invention divides images with high local similarity and difficult to classify into new composite classes, and uses the density of different classes to gradually merge the composite classes and single classes together with a given rule until satisfying the The number of new single classes after merging just stops when the actual number of image classifications, effectively reducing the risk of wrongly dividing the data, and avoiding making wrong or biased decisions.

【Description of drawings】

Fig. 1 is a flowchart of the present invention;

FIG. 2 is an example diagram of a target sample _xi and its K ₂ neighbor vectors in an embodiment of the present invention.

【Detailed ways】

The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

Asymmetric data clustering is a very challenging problem. The present invention proposes a new trust clustering method (CClu) for processing asymmetric data under the framework of trust function theory. For an image set containing c-type images, first use the trusted c-means clustering algorithm (CCM) to divide the image set into categories greater than c (the number of true categories) (for example, the number of categories is N, N>>c), These classes contain both the initial single class and the initial composite class, where the composite class is composed of images in overlapping inseparable regions of different classes, ie images like one class and another. In this way, images originally belonging to the same category can be clustered into the same category, and then the N categories can be gradually merged together with a given rule by using the density of different categories, until the combined new single category and the initial single category are satisfied. Stop when the sum is exactly c. Then, a vector method with weights is designed using the K-nearest neighbors of the image to re-classify those initial composite classes that have not been merged into new single classes or new composite classes.

Images that are difficult to classify in overlapping regions will be classified into new composite classes, in order to reduce the risk of misclassification. Multiple experiments are conducted with artificial and real datasets to test the performance of CClu relative to other related methods. The results show that CClu can handle multi-class asymmetric image sets well, and due to the existence of composite classes, the error is effectively reduced by correct modeling of sample imprecision.

The new CClu method is used to deal with asymmetric images under the framework of trust function theory. For a class c problem, CClu first uses CCM to generate several initial subclasses (including N initial single classes and several composite classes) that are more than the real class c of the image, where only two initial composite classes are allowed to contain single class. Then, use the K-nearest neighbors of the images contained in these initial subclasses to calculate the density of these initial subclasses, and use the given density relationship-based merging rule to merge the initial composite class and its initial single class into a New single class until the number of new single class is exactly c.

Finally, the images in the original composite class that were not merged are classified into the new single class or the new composite class using the K-nearest neighbors in the new single class. For images that are difficult to be accurately classified in the initial composite class, CClu will classify them into a new composite class consisting of a new single class, which will effectively reduce the error rate of image classification. Images in the new composite class are difficult to classify accurately under current conditions, but they can eventually be classified by new techniques or additional information. Therefore, CClu prevents bad decisions by generating special clustering results, which is important at all times.

An image classification method based on the trust function theory of the present invention, as shown in Figure 1, includes the following steps:

The c-means clustering algorithm is used to divide all images in the image set X to be classified into multiple initial subclasses; wherein the image set X to be classified is an image set of asymmetric data. The initial subclass consists of N initial single classes and multiple initial composite classes, each initial composite class consists of two initial single classes, N ≥ c, c represents the number of true classes of all images in the image set X.

个类别，其中，2c表示初始单类的数量，

表示初始复合类的数量，1表示噪声类φ。这些类别可以用幂集2^Ω的一个子集S^Ω表示，其中，S^Ω被定义为

t_c表示允许初始复合类所包含初始单类个数的阈值。定义初始单类ω_i和ω_j的复合类用A_i,j表示，所以当N＝2c时，

举个例子，对于一个c＝2类的问题，当N＝2c＝4时，Ω＝{ω₁,ω₂,ω₃,ω₄}，CCM在2^Ω框架中将会至多产生

个图像类别，即

Consider an image set containing c categories of images, under the guidance of multi-clustering, all images in the image set X will be clustered into different categories by CCM. For example, when N=2c, 2 ^2c classes will be generated, including the initial single class, the initial composite class and the noise class, which will bring a lot of computational problems as N increases. However, in practical applications, it is rare that the image is in the overlapping area of different categories with more than 2 initial single categories. Indistinguishable between initial single classes. Therefore, in the present invention, only the initial composite class is allowed to contain 2 initial single classes. At this time, when N=2c, CCM will at most generate

categories, where 2c represents the number of initial single categories,

represents the number of initial composite classes, and 1 represents the noise class φ. These classes can be represented by a subset S ^Ω of the power set 2 ^Ω , where S ^Ω is defined as

t _c represents the threshold for the number of initial single classes that are allowed to be included in the initial composite class. The composite class that defines the initial single class ω _i and ω _j is represented by A _i,j , so when N=2c,

For example, for a problem of class c = 2, when N = 2c = 4, Ω = {ω ₁ , ω ₂ , ω ₃ , ω ₄ }, CCM in the 2 ^Ω frame will yield at most

image categories, i.e.

Calculate the density value for each initial subclass. where density is the reciprocal of the mean Euclidean distance of all images in the initial subclass. In order to avoid accidental errors, the average Euclidean distance of the K ₁ nearest neighbors of the image x _i ∈ A _j is used to represent the nearest neighbor distance of x _i ∈ A _j , and the density of each initial subclass is calculated using the formula:

表示初始子类A_j中的图像x_i与其在图像集X中的K₁个近邻的欧氏距离的平均距离，

d_ik表示图像x_i及其在图像集中第k个近邻x_k之间的欧氏距离，K₁为常数，表示图像x_i的近邻个数，n_j为初始子类A_j中图像的数量，i为初始子类A_j中图像的序数。where ρ _j is the density of the j-th initial subclass A _j ,

represents the average distance of the Euclidean distance between the image x _i in the initial subclass A _j and its K ₁ nearest neighbors in the image set X,

d _ik represents the Euclidean distance between image _xi and its k-th nearest neighbor x _k in the image set, K ₁ is a constant, represents the number of neighbors of image _xi , n _j is the number of images in the initial subclass A _j , i is the ordinal number of images in the initial subclass A _j .

It can be found from the above formula that the more dispersed the objects in the category A _j , the smaller the value of ρ _j . ρ _j is used to describe the discrete degree of images in category A _j , so it can be used as the basic basis for making relevant judgments on category A _j .

These initial subclasses can be combined using the density relationship between the initial composite class and the initial single classes contained in the initial composite class. In the present invention, the initial composite class is regarded as a transition class between different initial single classes, and the images in the initial composite class are regarded as belonging to a certain initial single class contained in it, which is an inexact representation.

If the images originally belonging to a certain category are divided into different categories (including the initial single category and the initial composite category) due to multi-clustering, then these initial single categories should have little difference in density, and the initial composite The density of the class should be greater than or equal to the density of the initial single class it contains, because the images in the initial composite class should be distributed in the relative center of this class. It should be noted that not all initial composite classes satisfy the above principles, because in some special cases, the images in the initial composite class may be very few or no initial composite class, which means that the distribution of this data is rather special.

Therefore, according to the size relationship between the initial composite class and the density value of the corresponding initial single class, the images in the initial composite class and the corresponding initial single class are merged to generate a new single class and/or a new composite class, until a new single class is generated after merging The sum of the number of the initial single class and the initial single class is c, and the size relationship between the initial composite class and the corresponding initial single class density value consists of the following three cases:

Case 1: The density of the initial composite class is greater than or equal to the density of the initial single class it contains, that is, ρ _k,t ≥ρ _k and ρ _k,t ≥ρ _t .

Case 2: The density of the initial composite class is between the densities of the initial single class it contains, that is, ρ _k ≤ρ _k,t ≤ρ _t .

Case 3: The density of the initial composite class is smaller than the density of the initial single class it contains, ie ρ _k,t ≤ρ _k and ρ _k,t ≤ρ _t .

Among them, ρ _k is the density value of the k-th initial single class, 0≤k≤N, ρ _t is the density value of the t-th initial single class, 0≤t≤N, ρ _k,t is the k-th initial single class The density value of the initial composite class consisting of a single class and the t-th initial single class.

According to the above principles, the initial single class and the initial composite class that satisfy the first case should be merged into a new single class first, followed by the one satisfying the second case, and finally the third case, that is, the initial composite class and the corresponding initial single class in the The order of merging in image merging is as follows: Case 1, Case 2, Case 3. The new single class is defined as:

其中，D_d是密度距离集，R_(ρi)是ρ_i在集合D_s中所处位置的序号值，并且集合D_s表示的是不同初始单类的密度值按照大小降序排列的集合，被定义为D_s＝sort{ρ₁,...,ρ_c,...,ρ_2c|1≤i≤2c}，集合D_o表示不同初始复合类密度按照大小降序排列的集合，被定义为D_s＝sort{ρ_1,2,...,ρ_1,2c,...,ρ_2c-1,2c|1≤i,j≤2c}，在同一种情况下，当密度距离一致时集合D_o将会被用来辅助判断合并顺序的优先级。y_i,j是

和

之间差值的绝对值，即为密度距离。In a specific problem, it is possible to analyze only one of them, or it may be necessary to analyze two of them, or even all three of them need to be analyzed at the same time. If there are too many initial single classes and initial composite classes that satisfy the same situation (such as meeting the first case), the priority issue needs to be considered when merging. In order to avoid the confusion of choices that may occur when merging in the same situation, the present invention designs a priority rule,

Among them, D _d is the density distance set, R _(ρi) is the serial number value of the position of ρ _i in the set D _s , and the set D _s represents the set of the density values of different initial single classes arranged in descending order of size, which is Defined as D _s =sort{ρ ₁ ,...,ρ _c ,...,ρ _2c |1≤i≤2c}, the set D _o represents a collection of different initial composite class densities arranged in descending order of size, and is defined as D _s =sort{ρ _1,2 ,...,ρ _1,2c ,...,ρ _2c-1,2c |1≤i,j≤2c}, in the same case, when the density distances are consistent The set D _o will be used to help determine the priority of the merge order. y _i,j is

and

The absolute value of the difference between them is the density distance.

When the density distance between the two initial single classes included in the initial composite class is small, CClu will preferentially merge the initial composite class and the two initial single classes it contains into a new single class, which means that in the same case, The closer the density values of the two initial single classes contained in the initial composite class are (that is, the smaller the density distance), the more likely they are to be merged preferentially (in the same case, the density distance is used as the priority judgment criterion, because the density value is more It is easily affected by data distribution or clustering results, so its variation is very large). The purpose of this is not only to prevent the initial single class with large density difference from being merged into a new single class, but also to prevent these classes from being merged into a new single class when the density value of the initial composite class is larger due to partial data overlap. However, in the same case, when the density distances of the initial single classes included in the initial composite class are consistent, the initial composite class with a larger density value and the initial single class included in it will be merged preferentially.

The specific method of merging the images in the initial composite class and the corresponding initial single class is: calculating the density distance value between the two initial single classes corresponding to the initial composite class, and sorting the initial composite classes according to the density distance value from small to large, And sequentially merge the images in the initial composite class and the corresponding initial single class into a new single class.

When the density distance values between the two initial single classes corresponding to the initial composite class are equal, the images in the initial composite class and the corresponding initial single class are merged into one according to the order of the density value of each initial composite class from large to small. New single class. Among them, the density distance value is the difference between the ordinal numbers of the two initial single classes in the initial single class density set D _s , the initial single class density set D _s is the set of density values of N initial single class, and the initial single class density The set D _s is sorted according to the density value of the initial single class from large to small.

其中，

表示由合并法则传递性产生的新单类，

和

分别表示不同的新单类。After the new single class is generated and before the next initial composite class merging, it is judged whether the new single class and other new single classes have the same image. All images in are merged to generate a new single class. That is to say, in many cases, after merging different categories to generate a new single category, these new single categories may continue to be merged. That is to say, when two new single classes are generated, priority must be given to whether these two new single classes need to be merged again, and the condition for merging is that the intersection of these two new single classes is not an empty set, which is called is the transitivity of the merger rule, that is

in,

represents a new single class resulting from the transitivity of the merge rule,

and

respectively represent different new single classes.

A simple example is given below to further illustrate the merge rule.

Case 1: Consider a class 3 problem and assume N=2c=6, Ω={ω ₁ ,ω ₂ ,ω ₃ ,ω ₄ ,ω ₅ ,ω ₆ },S ^Ω ={ω ₁ ,ω ₂ ,ω ₃ ,ω ₄ ,ω ₅ ,ω ₆ ,ω _1,2 ,ω _1,3 ,ω _1,4 ,ω _1,5 ,ω _2,4 ,ω _2,6 ,ω _4,5 ,ω _4,6 }, in ^HΩ , ρ ₁ =2.17, ρ ₂ =3.2, ρ ₃ =1.85, ρ ₄ =1.59, ρ ₅ =1.05, ρ ₆ =1.23, ρ _1,2 =2.77, ρ _1,3 =2.02 , ρ _1,4 =2.26, ρ _1,5 =1.45, ρ _2,4 =2.66, ρ _2,6 =1.97, ρ _4,5 =0.9, ρ _4,6 =0.99, it is easy to find that the three cases correspond to There are different initial composite classes and the initial single class they contain, and can be obtained by calculation:

Satisfy condition 1: ρ _1,4 ;

Satisfy the second condition: ρ _1,2 , ρ _1,3 , ρ _1,5 , ρ _2,4 , ρ _2,6 ;

Satisfy condition three: ρ _4,5 , ρ _4,6 ;

Then, D _d ={y _1,2 =1,y _1,3 =1,y _1,4 =2,y _1,5 =5,y _2,4 =3,y _2,6 =4,y _4,5 =2,y _4,6 =1}, and there is, D _s ={ρ ₂ >ρ ₁ >ρ ₃ >ρ ₄ >ρ ₆ >ρ ₅ }, D _o ={ρ _1,2 >ρ _2,4 >ρ _1,4 >ρ _1,3 >ρ _2,6 >ρ _1,5 >ρ _4,6 >ρ _4,5 }.

According to the merging rule, the categories satisfying case one should be merged first, and only ρ _1,4 is satisfied. So ω _1,4 , ω ₁ and ω ₄ should be merged and produce a new single class. Now the number of new single classes does not satisfy N'=3, so it is necessary to continue to merge the classes that satisfy the second case.

There are five initial composite classes and the initial single class contained in them satisfy the second case, so the density distance should be introduced to assist in judging the priority merging order. Since y _1,2 =y _1,3 =1 is the smallest in the set D _d , ω _1,2 , ω _1,3 and the initial single class they contain should be merged preferentially, and their merge order is ω _1,2 first and then ω _1,3 , because in the set D _o ρ _1,2 >ρ _1,3 , the next merge order is ω _2,4 , ω _2,6 , ω _1,5 .

Obviously, among the initial composite classes and the initial single classes they contain that satisfy Case 3, the initial composite classes ω _4,6 and the initial single classes they contain should be merged first, because in the set D _d y _4,6 >y _4,5 .

It should be noted that although the merging order of all initial composite classes and the initial single classes they contain is analyzed in turn in three cases, not all analyses are useful, as can be seen from the following merging steps :

N′＞3。Step 1: In case 1,

N'>3.

然后可得，Step 2: In case 2,

then get,

且

N′＞3。

and

N'>3.

然后可得，Step 3: In case 2,

then get,

且

N′＝3。

and

N'=3.

Therefore, the new single class generated by the merge should be:

所以接下来需要把这4个初始复合类中的图像划分到新单类中或者划分到由这3个新单类组成的新的复合类中(例如

)，其中，

同样被认为是

和

组成的新复合类，处于这两个新单类的重叠区域。It should be noted that there are also 4 composite classes {ω _1,5 ,ω _2,6 ,ω _4,5 ,ω _4,6 } which are not merged into the new single class

Therefore, it is necessary to divide the images in these 4 initial composite classes into new single classes or into new composite classes composed of these 3 new single classes (for example

),in,

also considered to be

and

The new composite class formed is in the overlapping area of the two new single classes.

The images in the initial composite class are divided into new single class, initial single class or new composite class, and the classification result of the images in the image set X to be classified is obtained.

The images in the unmerged initial composite class will be divided into a new single class related to the original composite class in which it is located or a new composite class composed of these new single classes, that is, the images in the initial composite class will be divided into the new single class. class, initial single class, or new composite class. Therefore, a vector weighted cosine distance method is proposed based on the sample K nearest neighbors.

其中

表示x_j的类别是

例如，在案例1中复合类ω_1,5包含两个初始单类ω₁和ω₅，其中ω₁被合并到了新单类

中，并且ω₅被合并到了新单类

所以初始复合类ω_1,5中的图像x_i应该被划分到新单类

或者

中，又或者被划分到由它们组成的新复合类中

但是初始复合类ω_1,5中的图像x_i不能够被划分到新单类

或者与它相关的新复合类中，因此初始复合类ω_1,5中的图像x_i的K₂个近邻的标签只能够是

或者

即

或者

In this method, first, select the image _xi in the initial composite class, find the K ₂ nearest neighbors of the image _xi in the new single class, and define it as

in

The class representing x _j is

For example, in case 1 the composite class ω _1,5 contains two initial single classes ω ₁ and ω ₅ , where ω ₁ is merged into the new single class

, and ω ₅ is merged into the new single class

So the images _xi in the initial composite class ω _1,5 should be classified into the new single class

or

, or into a new composite class consisting of them

But the images _xi in the initial composite class ω _1,5 cannot be classified into the new single class

or in a new composite class related to it, so the labels of the K ₂ nearest neighbors of image x _i in the initial composite class ω _1,5 can only be

or

which is

or

的欧氏距离通常不相同，所以在使用向量法时需要采取加权策略，根据图像x_i和其近邻图像之间的欧氏距离生成K₂个向量的权重。一般而言，图像离它的近邻越远，那么近邻的可靠性就越低。所以距离d_ij越大往往导致其可靠性权重λ_ij就越低。一个简单又合理的方法已经被广泛应用于各个领域，在这里选择一种被广泛应用于各个领域且非常合理的估计可靠性权重的方法，

其中d_ij表示图像x_i和近邻

之间的欧氏距离。Select the image _xi in the initial composite class, and find the K ₂ images with the nearest Euclidean distance in the image set X to be classified as its nearest neighbor images, and generate K ₂ vectors. Since images x _i in the initial composite class are not merged to K ₂ nearest neighbors

The Euclidean distances are usually different, so a weighting strategy needs to be adopted when using the vector method, and the weights of K ₂ vectors are generated according to the Euclidean distance between the image _xi and its neighboring images. In general, the farther an image is from its neighbors, the less reliable the neighbors are. Therefore, the larger the distance d _ij , the lower the reliability weight λ _ij will be. A simple and reasonable method has been widely used in various fields. Here, a method that is widely used in various fields and very reasonable to estimate reliability weights is selected.

where d _ij represents the image _xi and the nearest neighbors

Euclidean distance between .

Calculate the sum vector of K ₂ vectors as the first sum vector;

Calculate the sum vector of the vectors of the vector end points corresponding to the vector end points in the K ₂ vectors that belong to the same initial single class/new single class, as the second sum vector;

Calculate the cosine value of the angle formed by each second sum vector and the first sum vector, and divide the image _xi into the initial single class/new single class corresponding to the second sum vector with the smallest cosine value of the included angle.

其中，

表示向量的起点是未合并初始复合类中的图像x_i，向量的终点是K₂个近邻中的近邻x_j，并且x_j向量加权余弦距离方法的核心思想是分别计算图像x_i的K₂个近邻中属于同一个类别的所有向量的和向量，并将它们同K₂个向量

的和向量作比较，余弦值较小的类别向量所代表的类别作为图像x_i的最终划分类别。也就是说，不同类的和向量与所有向量的和向量之间夹角的余弦越小，图像x_i就越有可能属于余弦值较小的类别。当不同类的和向量与所有向量的和向量之间夹角的余弦值非常接近(差异不大)时，这意味着图像x_i很难被分到某个新单类。也就是说，此时图像x_i应该被划分到由新单类组成的新复合类中。重复上述步骤，直至将所有初始复合类中的图像划分完。After getting K ₂ nearest neighbors, you can get K ₂ vectors

in,

The starting point of the representation vector is the image x _i in the unmerged initial composite class, the end point of the vector is the nearest neighbor x _j among the K ₂ nearest neighbors, and the core idea of the x _j vector weighted cosine distance method is to calculate the K ₂ of the image x _i separately The sum vector of all vectors belonging to the same category in the nearest neighbors, and compare them with K ₂ vectors

The sum vector is compared, and the category represented by the category vector with the smaller cosine value is used as the final classification category of the image _xi . That is, the smaller the cosine of the angle between the sum vector of different classes and the sum vector of all vectors, the more likely the image _xi belongs to the class with the smaller cosine value. When the cosine value of the angle between the sum vector of different classes and the sum vector of all vectors is very close (not very different), it means that the image _xi is difficult to be classified into a new single class. That is, at this point the image _xi should be classified into a new composite class consisting of a new single class. Repeat the above steps until all images in the initial composite class are divided.

When the difference between the cosine value of the angle formed by the two second sum vectors and the first sum vector is less than the threshold, divide the image into a new composite formed by the initial single class/new single class corresponding to the two second sum vectors in the class.

并且d_ik表示向量

的模或者图像x_i和近邻x_j之间欧氏距离。所以向量加权余弦距离方法的计算步骤如下：As shown in Figure 2, the calculation process of the vector weighted cosine distance method is briefly introduced. In the figure, assuming K ₂ =5,

and d _ik represents the vector

or the Euclidean distance between the image _xi and its nearest neighbors x _j . So the calculation steps of the vector weighted cosine distance method are as follows:

The first step: solve Γ, Γ ₁ , Γ ₂ ;

和

Step 2: Solve

and

Step 3: According to the initial composite class parameter ξ, the image _xi belongs to a new single class or a new composite class composed of the new single class.

Parameter setting rules: According to the experimental results, N=2c is recommended as the default value. If N is too small, the performance of CClu will decrease, but if the value of N is too large, many categories will be generated, which will bring a huge amount of calculation; ξ ∈ [0, 0.4] is also set as the default value, which should be set according to the accuracy acceptable to the user in the specific application process. At the same time, since K ₁ and K ₂ are between [5, 10], they have little effect on the experimental results. Therefore, the default values of K ₁ and K ₂ are set to 5.

来获取，其中，0≤N*(c)≤1。In CClu, it is assumed that the real category c can be derived from prior knowledge or experience. If no prior knowledge is available, it can be obtained by

to obtain, where 0≤N*(c)≤1.

Claims (9)

1. the image classification method based on trust function theory, is characterized in that, comprises the following steps: Divide all images in the image set X to be classified into multiple initial subclasses; wherein, the initial subclass consists of N initial single classes and multiple initial composite classes, and each initial composite class consists of two initial composite classes. Single-class composition, N≥c, c represents the number of real classes of all images in the image set X; Calculate the density value of each of the initial subclasses; wherein, the density is the reciprocal of the average Euclidean distance between all images in the initial subclass; The density of each of the initial subclasses is calculated using the formula:

where ρ _j is the density of the j-th initial subclass A _j ,

represents the average distance of the Euclidean distance between the image x _i in the initial subclass A _j and its K ₁ nearest neighbors in the image set X, where K ₁ is a constant, n _j is the number of images in the initial subclass A _j , and i is ordinal number of images in initial subclass A _j ; According to the size relationship between the initial composite class and the density value of the corresponding initial single class, the images in the initial composite class and the corresponding initial single class are merged to generate a new single class and a new composite class, until a new single class is generated after merging and the sum of the initial single class quantity is c; The images in the initial composite class are divided into the new single class, the initial single class or the new composite class, and the classification result of the images in the image set X to be classified is obtained. 2. The image classification method based on the trust function theory as claimed in claim 1, wherein the size relationship between the initial composite class and the corresponding initial single class density value is made up of the following three situations: Case 1: ρ _k,t ≥ρ _k and ρ _k,t ≥ρ _t ; Case 2: ρ _k ≤ρ _k,t ≤ρ _t ; Case 3: ρ _k,t ≤ρ _k and ρ _k,t ≤ρ _t ; Among them, ρ _k is the density value of the k-th initial single class, 0≤k≤N, ρ _t is the density value of the t-th initial single class, 0≤t≤N, ρ _k,t is the k-th initial single class The density value of the initial composite class consisting of a single class and the t-th initial single class. 3. The image classification method based on the trust function theory as claimed in claim 2, wherein the merging sequence when merging the images in the initial composite class and the corresponding initial single class is: case one, case two, Case three. 4. The image classification method based on the trust function theory as claimed in claim 3, wherein the concrete method for merging the images in the initial composite class and the corresponding initial single class is: Calculate the density distance value between the two initial single classes corresponding to the initial composite class, sort the initial composite classes according to the density distance value from small to large, and sequentially place the initial composite class and the corresponding initial single class in the The images are merged into a new single class; When the density distance values between the two initial single classes corresponding to the initial composite class are equal, the initial composite class and the corresponding initial single class are sorted in order according to the density value of each initial composite class in descending order. Images are merged into a new single class; Wherein, the density distance value is the difference between the ordinal numbers of the two initial single classes in the initial single class density set D _s , and the initial single class density set D _s is the density value of the N initial single class , and the initial single-class density set D _s is sorted according to the density value of the initial single-class from large to small. 5. The image classification method based on the trust function theory as claimed in claim 4, characterized in that, after a new single class is generated and before the next initial composite class merging is performed, it is judged that the new single class is different from other new single classes. Whether there is the same image, when it exists, combine the new single class with all the images in the new single class with the same image to generate a new single class. 6. The image classification method based on the trust function theory according to any one of claims 2-5, wherein the images in the initial composite class are divided into the new single class, the initial single class or the new composite class The specific methods are: Select the image x _i in the initial composite class, and find the K ₂ images with the closest Euclidean distance in the image set X to be classified as its nearest neighbor images, and generate K ₂ vectors, according to the image The Euclidean distance between x _i and its nearest neighbor images generates weights of K ₂ vectors; Calculate the sum vector of K ₂ vectors as the first sum vector; Calculate the sum vector of the vectors of the vector end points corresponding to the vector end points in the K ₂ vectors that belong to the same initial single class/new single class, as the second sum vector; Calculate the cosine value of the angle formed by each of the second sum vectors and the first sum vector, and divide the image x _i into the initial single class/new single class corresponding to the second sum vector with the smallest cosine value of the included angle ; Repeat the above steps until all images in the initial composite class are divided. 7. The image classification method based on the trust function theory as claimed in claim 6, wherein when the difference between the cosine values of the included angles formed by the two second sum vectors and the first sum vector is less than a threshold, the The image is divided into a new composite class formed by the initial single class/new single class corresponding to the two second sum vectors. 8 . The image classification method based on the trust function theory according to claim 7 , wherein all images in the image set X to be classified are divided into a plurality of initial sub-classes, and a c-means clustering algorithm is specifically adopted. 9 . 9 . The image classification method based on the trust function theory according to claim 7 or 8 , wherein the image set X to be classified is an image set of asymmetric data. 10 .

Images