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

Image classification method based on trust function theory Download PDF

Info

Publication number
CN110472657B
CN110472657B CN201910599618.7A CN201910599618A CN110472657B CN 110472657 B CN110472657 B CN 110472657B CN 201910599618 A CN201910599618 A CN 201910599618A CN 110472657 B CN110472657 B CN 110472657B
Authority
CN
China
Prior art keywords
initial
class
classes
images
image
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201910599618.7A
Other languages
Chinese (zh)
Other versions
CN110472657A (en
Inventor
刘准钆
张作伟
潘泉
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Northwestern Polytechnical University
Original Assignee
Northwestern Polytechnical University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Northwestern Polytechnical University filed Critical Northwestern Polytechnical University
Priority to CN201910599618.7A priority Critical patent/CN110472657B/en
Publication of CN110472657A publication Critical patent/CN110472657A/en
Application granted granted Critical
Publication of CN110472657B publication Critical patent/CN110472657B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering

Abstract

The invention discloses an image classification method based on a trust function theory, which divides all images in an image set X to be classified into a plurality of initial subclasses; calculating the density value of each initial subclass; combining the images in the initial composite class and the corresponding initial single class according to the magnitude relation of the density values of the initial composite class and the corresponding initial single class to generate a new single class and a new composite class until the sum of the number of the new single class and the number of the initial single class generated after combination is c; dividing the images in the initial composite class into a new single class, an initial single class or a new composite class to obtain the classification result of the images in the image set X to be classified; the invention gradually merges the composite classes and the single classes together by a given rule by utilizing the densities of the different classes until the initial number of the merged single classes is just equal to the real number of the classified images, thereby effectively reducing the risk brought by mistakenly dividing data and avoiding making wrong or more-deviated decisions.

Description

Image classification method based on trust function theory
[ technical field ] A method for producing a semiconductor device
The invention belongs to the technical field of image classification and identification, and particularly relates to an image classification method based on a trust function theory.
[ background of the invention ]
Unbalanced data clustering is one of the important branches of pattern recognition and unsupervised machine learning, and has wide application in various fields.
Image recognition of airborne targets is an important direction of research in aerospace monitoring systems. Due to the reasons of artificial countermeasure, complex and variable battlefield environment, limited sensor performance and the like, the image sensor can generate the condition of unbalanced quantity of images of different types of moving targets when shooting images of moving targets in the air, which is specifically represented by the fact that the shot images of different types of moving targets have large difference in quantity, and meanwhile, the images of different targets under certain visual angles are possibly very similar, so that the target images are obtained with great uncertainty.
When the existing classical clustering method carries out unsupervised classification (clustering) on the unbalanced samples, a reliable result is often difficult to obtain, and even a 'balance effect' occurs, namely the number of images of different targets is very different, for example, when images are shot for two targets at the same time, the number of images of one target is very small, the number of images of the other target is very large, but after clustering, an image set can still be clustered into two target categories with approximately equal number, so that a high error rate occurs during image classification. Meanwhile, the traditional classical clustering method does not consider the problem that images of different targets are likely to be very similar under certain visual angles, namely the images of the different targets are likely to be similar in part positions, and when the similarity of the images belonging to the different targets is very high, the overlapping images are difficult to be clustered accurately.
[ summary of the invention ]
The invention aims to provide an image classification method based on a trust function theory, which is used for classifying images by using the density of various images as a classification standard when the images of different targets are classified, so that the accuracy of image classification is improved.
The invention adopts the following technical scheme: the image classification method based on the trust function theory comprises the following steps:
dividing all images in an image set X to be classified into a plurality of initial subclasses; the initial sub-class consists of N initial single classes and a plurality of initial composite classes, each initial composite class consists of two initial single classes, N is larger than or equal to c, and c represents the real class number of all images in the image set X;
calculating the density value of each initial subclass; wherein the density is the reciprocal of the mean value of Euclidean distances of all images in the initial subclass;
combining the images in the initial composite class and the corresponding initial single class according to the magnitude relation of the density values of the initial composite class and the corresponding initial single class to generate a new single class and a new composite class until the sum of the number of the new single class and the number of the initial single class generated after combination is c;
and dividing the images in the initial composite class into a new single class, an initial single class or a new composite class to obtain the classification result of the images in the image set X to be classified.
Further, the magnitude relationship between the initial composite class and the corresponding initial single class density value is composed of the following three conditions:
the first condition is as follows: rhok,t≥ρkAnd ρk,t≥ρt；
Case two: rhok≤ρk,t≤ρt；
Case three: rhok,t≤ρkAnd ρk,t≤ρt；
Where ρ iskK is the density value of the kth initial single class, k is more than or equal to 0 and less than or equal to N, rhotIs the density value of the t initial single class, t is more than or equal to 0 and less than or equal to N, rhok,tIs the density value of the initial composite class consisting of the kth initial single class and the tth initial single class.
Further, the merging sequence when merging the images in the initial composite class and the corresponding initial single class sequentially is as follows: case one, case two, case three.
Further, the specific method for merging the images in the initial composite class and the corresponding initial single class is as follows:
calculating a density distance value between two initial single classes corresponding to the initial composite class, sequencing the initial composite classes according to the density distance values from small to large, and combining the images in the initial composite class and the images in the corresponding initial single classes into a new single class in sequence;
when the density distance values between two initial single classes corresponding to the initial composite classes are equal, combining the images in the initial composite classes and the images in the corresponding initial single classes into a new single class according to the descending order of the density values of the initial composite classes;
wherein the density distance value is two initial single-class density sets D in the initial single-classsDifference of ordinal number in (1), initial single class density set DsIs a set of density values of N initial single classes, and an initial single class density set DsThe density values of the initial single classes are sorted from large to small.
Further, after the new single class is generated and before the next initial composite class combination, whether the new single class and other new single classes have the same images or not is judged, and if the new single class and all the images in the new single class with the same images exist are combined to generate a new single class.
Further, the specific method for classifying the images in the initial composite class into the new single class, the initial single class or the new composite class is as follows:
selecting an image x in an initial composite classiAnd finding out the K closest to the Euclidean distance from the image set X to be classified2The individual image is used as its neighbor image, and K is generated2Vector of quantity, from image xiAnd its adjacent image Euclidean distance between generate K2A weight of the vector;
calculate K2A sum vector of the vectors as a first sum vector;
calculate K2The adjacent image corresponding to the vector end point in the vector belongs to the sum vector of the vectors of the same initial single type/new single type, and the sum vector is used as a second sum vector;
calculating cosine value of included angle formed by each second sum vector and the first sum vector, and displaying the image xiDividing the initial single type/new single type corresponding to the second sum vector with the minimum cosine value of the included angle;
and repeating the steps until the images in all the initial compound classes are divided.
Further, when the difference value of cosine values of included angles formed by the two second sum vectors and the first sum vector is smaller than a threshold value, dividing the image into a new compound class formed by the initial single class/new single class corresponding to the two second sum vectors.
Further, all images in the image set X to be classified are divided into a plurality of initial subclasses, and a 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 by the following formula:
Figure BDA0002118756180000041
where ρ isjIs the jth initial subclass AjThe density of (a) of (b),
Figure BDA0002118756180000042
represents the initial subclass AjImage x of (1)iWith its K in image set X1Mean distance of Euclidean distances of neighbors, K1Is a constant number, njIs an initial subclass AjNumber of middle images, i is the initial subclass AjThe ordinal number of the image in (1).
The invention has the beneficial effects that: the invention divides the images with high local similarity and difficult classification into new compound classes, gradually combines the compound classes and the single classes together by a given rule by utilizing the density of the different classes until the number of the new single classes after combination is just equal to the number of the real classification of the images, effectively reduces the risk brought by mistakenly dividing data and avoids making wrong or more-wrong decisions.
[ description of the drawings ]
FIG. 1 is a block flow diagram of the present invention;
FIG. 2 shows a target sample x according to an embodiment of the present inventioniAnd K thereof2Example graph of neighbor vectors.
[ detailed description ] embodiments
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, and the invention provides a new trust clustering method (CClu) for processing asymmetric data under the trust function theory framework. For an image set containing c types of images, firstly, the image set is divided into categories (for example, the number of categories is N, N > > c) which are larger than c (the number of real categories) by using a trusted c-means clustering algorithm (CCM), wherein the categories comprise an initial single category and an initial composite category, and the composite category is composed of images in different categories of overlapping non-partitioned areas, namely, the images are similar to the category and the other category. Thus, images originally belonging to the same category can be clustered into the same category, and then the N categories are gradually merged together by a given rule by using the densities of different categories until the sum of the new and initial categories after merging is just c. Then, a vector method with weight is designed by utilizing the K neighbors of the image to re-classify the initial composite classes which are not combined into a new single class or a new composite class.
Images which are difficult to classify in the overlapping area are classified into new compound classes, and the purpose of this is to reduce the risk of misclassification. Multiple experiments were performed using the artificial and real datasets to test the performance of CClu relative to other related methods. The results show that CClu can handle multiple classes of asymmetric image sets well, and errors are effectively reduced by correct modeling of sample inaccuracies due to the presence of complex classes.
The new CClu method is used to process asymmetric images under the trust function theory framework. For a class c problem, CClu first generates several initial sub-classes (containing N initial single classes and several composite classes) with CCM more than the true class c of the image, where only two initial single classes are allowed to be contained in the initial composite class. Then, the density of the initial sub-classes is calculated by using the K neighbors of the images contained in the initial sub-classes, and the initial composite classes and the initial single classes contained in the initial composite classes are merged into a new single class by using a given merging rule based on density relation until the number of the new single classes is just c.
And finally, dividing the images in the initial composite class which are not merged into the new single class or the new composite class by using the K neighbors in the new single class. For images in the initial compound class which are difficult to accurately classify, CClu classifies the images into a new compound class consisting of new single classes, which effectively reduces the error rate of image classification. Images in new compound classes are difficult to accurately partition under current conditions, but they can eventually be partitioned by new techniques or additional information. Therefore, it is important at any time that CClu prevents erroneous decisions from being made by generating special cluster results.
The image classification method based on the trust function theory, as shown in fig. 1, comprises the following steps:
dividing all images in the image set X to be classified into a plurality of initial subclasses by adopting a c-means clustering algorithm; wherein, the image set X to be classified is an image set of asymmetric data. The initial sub-class consists of N initial single classes and a plurality of initial compound classes, each initial compound class consists of two initial single classes, N is larger than or equal to c, and c represents the real class number of all the images in the image set X.
Considering that an image set contains c classes of images, all the images in the image set X are clustered into different classes by CCM under the guiding idea of multi-clustering. For example, when N is 2c, 2 will be generated2cThe individual classes, including the initial single class, the initial composite class, and the noise class, present a significant computational problem as N increases. However, in practical applications, it is rare that the image is in an overlapping area of different categories exceeding 2 initial single categories, that is, when a certain sample is difficult to accurately classify into a certain initial single category, it is difficult to distinguish between 2 initial single categories to a large extent. In the present invention, therefore, only the initial composite class is allowed to contain 2 initial single classes. At most CCM is generated when N is 2c
Figure BDA0002118756180000061
A category, where 2c represents the number of initial singletons,
Figure BDA0002118756180000062
indicating the number of initial complex classes and 1 the noise class phi. These categories may be grouped by power set 2ΩA subset S ofΩIs shown in which SΩIs defined as
Figure BDA0002118756180000071
tcA threshold value representing the number of initial single classes that the initial composite class is allowed to contain. Defining an initial single class ωiAnd ωjA for a complex ofi,jSo that when N is 2c,
Figure BDA0002118756180000077
for example, for a problem of class c-2, when N-2 c-4, Ω ═ ω ═ 41,ω2,ω3,ω4CCM at 2ΩWill at most generate in the frame
Figure BDA0002118756180000072
An image class, i.e.
Figure BDA0002118756180000073
The density value of each initial subclass is calculated. Wherein the density is the inverse of the mean of the euclidean distances of all images in the initial sub-class. To avoid accidental errors, first image x is usedi∈AjK of1Mean Euclidean distance of neighbors to represent xi∈AjThe density of each initial subclass is calculated by adopting the formula:
Figure BDA0002118756180000074
where ρ isjIs the jth initial subclass AjThe density of (a) of (b),
Figure BDA0002118756180000075
represents the initial subclass AjImage x of (1)iWith its K in image set X1The average distance of the euclidean distances of the neighbors,
Figure BDA0002118756180000076
dikrepresenting an image xiAnd its k-th neighbor x in the image setkBetween them, K1Is constant, represents the image xiN is the number of neighbors ofjIs an initial subclass AjNumber of middle images, i is the initial subclass AjThe ordinal number of the image in (1).
As can be seen from the above formula, class AjThe more dispersed the target in (1), ρjThe smaller the value of (c). RhojIs used to describe class AjOf the degree of dispersion of the image, so it can be used as a measure for class AjAnd carrying out relevant judgment according to the basic basis.
The initial compound classes may be merged using density relationships between the initial compound classes and the initial single classes contained within the initial compound classes. In the present invention, an initial compound class is considered as a transition class between different initial single classes, and an image in the initial compound class is considered to belong to a certain initial single class contained in the initial compound class, and is an inaccurate representation.
If images originally belonging to a class are classified into different classes (including initial single classes and initial composite classes) due to multi-clustering, the initial single classes should be not very different in density magnitude, and the density of the initial composite 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 originally be distributed relatively centrally in this class. It should be noted that not all of the initial composite classes satisfy the above principle, since there are some special cases where there may be very few or no images in the initial composite classes, which means that the distribution of the data is relatively special.
Therefore, according to the magnitude relation between the initial composite class and the density value of the corresponding initial single class, 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 the sum of the number of the new single class and the initial single class generated after merging is c, and the magnitude relation between the initial composite class and the density value of the corresponding initial single class is composed of the following three conditions:
the first condition is as follows: the density of the initial composite class is greater than or equal to the density of the initial single class it contains, i.e. ρk,t≥ρkAnd ρk,t≥ρt。
Case two: the density of the initial composite class is between the densities of the initial single classes it contains, i.e. pk≤ρk,t≤ρt。
Case three: the density of the initial composite class is less than the density of the initial single class it contains, i.e. pk,t≤ρkAnd ρk,t≤ρt。
Where ρ iskK is the density value of the kth initial single class, k is more than or equal to 0 and less than or equal to N, rhotIs the density value of the t initial single class, t is more than or equal to 0 and less than or equal to N, rhok,tFor an initial composite class composed of the kth initial single class and the tth initial single classThe density value of (a).
According to the above principle, the initial single class and the initial composite class satisfying the condition one should be merged into a new single class firstly, secondly satisfying the condition two, and finally satisfying the condition three, that is, the merging sequence when the images in the initial composite class and the corresponding initial single class are merged is sequentially: case one, case two, case three. The new singleton class is defined as:
Figure BDA0002118756180000091
in a particular problem, it is possible to analyze only one of the cases, two of them, or even all three cases simultaneously. If too many initial single classes and initial composite classes satisfy the same condition (e.g., condition one), then priority considerations may also need to be taken into account in merging at this time. In order to avoid the possible confusion of choices when merging in the same situation, the present invention designs a priority rule,
Figure BDA0002118756180000092
wherein D isdIs the set of density distances, R(ρi)Is ρiIn set DsThe sequence number of the position in (1), and set DsRepresenting a set of density values of different initial classes in descending order of their size, defined as Ds＝sort{ρ1,...,ρc,...,ρ2cI is more than or equal to 1 and less than or equal to 2c, and set DoA set representing different initial composite class densities in descending order of size is defined as Ds＝sort{ρ1,2,...,ρ1,2c,...,ρ2c-1,2cI is less than or equal to 1, j is less than or equal to 2c, and under the same condition, when the density distances are consistent, the set D is collectedoWill be used to assist in determining the priority of the merge order. y isi,jIs that
Figure BDA0002118756180000093
And
Figure BDA0002118756180000094
difference between themThe absolute value of the value is the density distance.
When the density distance between two initial single classes contained in the initial composite class is small, CClu preferentially merges the initial composite class and the two initial single classes contained in the initial composite class into a new single class, which means that the closer the density values of the two initial single classes contained in the initial composite class are in the same case (i.e. the smaller the density distance), the more likely they are to be preferentially merged (in the same case, the density distance is used as a priority criterion, because the density value is more easily affected by the data distribution or the clustering result, so its variation is very large). The purpose of this is to prevent the original single classes with large density difference from being merged into a new single class; and simultaneously, the classes are prevented from being merged into a new single class when the density value of the initial composite class is larger due to partial data overlapping. In the same case, however, when the density distances of the initial single classes contained in the initial composite class are consistent, the initial composite class with a higher density value and the initial single classes contained in the initial composite class are merged preferentially.
The specific method for merging the images in the initial composite class and the corresponding initial single class comprises the following steps: and calculating a density distance value between two initial single classes corresponding to the initial composite class, sequencing the initial composite classes from small to large according to the density distance value, and combining the images in the initial composite class and the images in the corresponding initial single classes into a new single class in sequence.
When the density distance values between two initial single classes corresponding to the initial composite classes are equal, images in the initial composite classes and the images in the corresponding initial single classes are combined into a new single class according to the descending order of the density values of the initial composite classes. Wherein the density distance value is two initial single-class density sets D in the initial single-classsDifference of ordinal number in (1), initial single class density set DsIs a set of density values of N initial single classes, and an initial single class density set DsThe density values of the initial single classes are sorted from large to small.
After the new single type is generated and before the next initial composite type combination, whether the same image exists in the new single type and other new single types is judged, and if so, whether the same image exists in the new single type and other new single types is judgedAnd merging all the images in the new single class and the new single class with the same images to generate a new single class. That is, in many cases, after merging different categories to produce a new single category, the new single categories may continue to be merged. That is, after two new single classes are generated, the two new single classes need to be considered preferentially without being merged again, and the condition of merging is that the intersection of the two new single classes is not an empty set, which is called the transitivity of the merging rule, i.e. the transitivity of the merging rule
Figure BDA0002118756180000101
Wherein the content of the first and second substances,
Figure BDA0002118756180000102
representing a new single class transitively generated by the merge rule,
Figure BDA0002118756180000103
and
Figure BDA0002118756180000104
respectively, representing different new single classes.
The merge rule is further illustrated below by way of a simple example.
Case 1: consider a class 3 problem and assume N2 c 6 and Ω ω1,ω2,ω3,ω4,ω5,ω6}，SΩ＝{ω1,ω2,ω3,ω4,ω5,ω6,ω1,2,ω1,3,ω1,4,ω1,5,ω2,4,ω2,6,ω4,5,ω4,6At HΩIn ρ1＝2.17，ρ2＝3.2，ρ3＝1.85，ρ4＝1.59，ρ5＝1.05，ρ6＝1.23，ρ1,2＝2.77，ρ1,3＝2.02，ρ1,4＝2.26，ρ1,5＝1.45，ρ2,4＝2.66，ρ2,6＝1.97，ρ4,5＝0.9，ρ4,6At 0.99, it is easy to find that the three cases correspond to different initial composite classes andthe initial single classes it contains, and can be calculated:
the following conditions are met: rho1,4；
The second condition is satisfied: rho1,2，ρ1,3，ρ1,5，ρ2,4，ρ2,6；
The third condition is satisfied: rho4,5，ρ4,6；
Then there is, Dd＝{y1,2＝1,y1,3＝1,y1,4＝2,y1,5＝5,y2,4＝3,y2,6＝4,y4,5＝2,y 4,61, and has, Ds＝{ρ2＞ρ1＞ρ3＞ρ4＞ρ6＞ρ5}，Do＝{ρ1,2＞ρ2,4＞ρ1,4＞ρ1,3＞ρ2,6＞ρ1,5＞ρ4,6＞ρ4,5}。
According to the merging rule, classes that satisfy case one should be merged first, and only ρ1,4And (4) meeting the requirement. So omega1,4，ω1And ω4Should be merged and a new single class is generated. Now the number of new singletons does not satisfy N' ═ 3, so it is necessary to continue merging the classes that satisfy case two.
There are five initial composite classes and the initial single classes contained therein satisfy case two, so the density distance should be introduced to assist in determining the preferential merging order. Due to the fact that in set DdMiddle y1,2＝y 1,31 is the smallest, so ω1,2、ω1,3And the initial single classes they contain should be merged first, and their merging order is ω first1,2Rear omega1,3Because in set DoMiddle rho1,2＞ρ1,3The next merging order is ω2,4，ω2,6，ω1,5。
Obviously, in the initial composite class satisfying case three and the initial single class included therein, the initial composite class ω is4,6And the initial single classes it contains should be merged first, since in set DdMiddle y4,6＞y4,5。
It should be noted that although the merging order of all the initial composite classes and the initial single classes contained therein is analyzed in turn in three cases, not all the analysis is useful, as can be seen from the following merging steps:
the method comprises the following steps: in the case of the first embodiment of the present invention,
Figure BDA0002118756180000111
N′＞3。
step two: in the case of the second situation, the user can,
Figure BDA0002118756180000112
and then it is possible to obtain,
Figure BDA0002118756180000113
and is
Figure BDA0002118756180000114
N′＞3。
Step three: in the case of the second situation, the user can,
Figure BDA0002118756180000115
and then it is possible to obtain,
Figure BDA0002118756180000116
and is
Figure BDA0002118756180000117
N′＝3。
Therefore, the new single class generated by the merging should be:
Figure BDA0002118756180000121
Figure BDA0002118756180000122
Figure BDA0002118756180000123
note that there are 4 more complex classes ω1,5,ω2,6,ω4,5,ω4,6Are not merged into a new single class
Figure BDA0002118756180000124
It is then necessary to classify the images in these 4 initial compound classes into new singlets or into a new compound class consisting of these 3 new singlets (for example
Figure BDA0002118756180000125
) Wherein, in the step (A),
Figure BDA0002118756180000126
are also considered to be
Figure BDA0002118756180000127
And
Figure BDA0002118756180000128
and forming a new composite class in the overlapping area of the two new single classes.
And dividing the images in the initial composite class into a new single class, an initial single class or a new composite class to obtain the classification result of the images in the image set X to be classified.
The images in the un-merged initial compound class will be classified into new single classes associated with the initial compound class in which they are located or new compound classes composed of these new single classes, i.e., the images in the initial compound class are classified into new single classes, initial single classes or new compound classes. Therefore, a vector weighted cosine distance method is proposed based on the sample K neighbor.
In the method, first, an image x in an initial composite class is selectediFinding an image x in a new document classiK of2Is adjacent and is defined as
Figure BDA0002118756180000129
Wherein
Figure BDA00021187561800001210
Denotes xjIs of the class
Figure BDA00021187561800001211
For example, in case 1 the complex class ω1,5Comprising two initial single classes ω1And ω5Wherein ω is1Is merged into a new single class
Figure BDA00021187561800001212
And ω is5Is merged into a new single class
Figure BDA00021187561800001213
So the initial complex class ω1,5Image x of (1)iShould be divided into new single classes
Figure BDA00021187561800001214
Or
Figure BDA00021187561800001215
Or divided into new compound classes composed of them
Figure BDA00021187561800001216
But the initial complex class omega1,5Image x of (1)iCannot be classified into new single classes
Figure BDA00021187561800001217
Or a new complex class associated with it, thus the initial complex class omega1,5Image x of (1)iK of2The labels of the neighbors can only be
Figure BDA00021187561800001218
Or
Figure BDA00021187561800001219
Namely, it is
Figure BDA00021187561800001220
Or
Figure BDA00021187561800001221
Selecting an image x in an initial composite classiAnd finding out the K closest to the Euclidean distance from the image set X to be classified2The individual image is used as its neighbor image, and K is generated2A vector. Due to uncombined image x in the initial composite classiTo K2One near neighbor
Figure BDA0002118756180000131
The Euclidean distances are usually different, so a weighting strategy needs to be adopted when a vector method is used, and the image x is obtainediAnd its adjacent image Euclidean distance between generate K2The weight of each vector. In general, the farther an image is from its neighbors, the lower the reliability of the neighbors. So that the distance dijLarger values tend to result in a reliability weight λ thereofijThe lower. A simple and reasonable method has been widely applied to various fields, where a method of estimating reliability weights that is widely applied to various fields and is very reasonable is selected,
Figure BDA0002118756180000132
wherein d isijRepresenting an image xiAnd neighbor
Figure BDA0002118756180000133
The euclidean distance between.
Calculate K2A sum vector of the vectors as a first sum vector;
calculate K2The adjacent image corresponding to the vector end point in the vector belongs to the sum vector of the vectors of the same initial single type/new single type, and the sum vector is used as a second sum vector;
calculating cosine value of included angle formed by each second sum vector and the first sum vector, and displaying the image xiDividing to the second sum vector with the minimum cosine value of the included angleCorresponding initial/new single classes.
In obtaining K2K can be obtained after one neighbor2An individual vector
Figure BDA0002118756180000134
Wherein the content of the first and second substances,
Figure BDA0002118756180000135
representing a starting point of the vector as image x in the initial composite class that is not mergediThe end point of the vector is K2Neighbor x of neighborjAnd x isjThe core idea of the vector weighted cosine distance method is to calculate the image x respectivelyiK of2All vectors belonging to the same class in a neighbor are summed and matched with K2An individual vector
Figure BDA0002118756180000136
The sum vector of (2) is compared, and the class represented by the class vector with smaller cosine value is taken as the image xiTo the final classification category. That is, the smaller the cosine of the angle between the sum vector of the different classes and the sum vector of all vectors, the image xiThe more likely it is to belong to a category with smaller cosine values. When the cosine values of the angles between the sum vectors of different classes and the sum vectors of all vectors are very close (not much different), this means that image x is an image of which the cosine values are not very closeiIt is difficult to be classified into a new single class. That is, image x at this timeiShould be divided into new composite classes consisting of new single classes. And repeating the steps until the images in all the initial compound classes are divided.
And when the difference value of cosine values of included angles formed by the two second sum vectors and the first sum vector is smaller than a threshold value, dividing the image into a new compound class formed by the initial single class/new single class corresponding to the two second sum vectors.
As shown in fig. 2, the calculation process of the vector weighted cosine distance method is briefly introduced. In the figure, assume K2＝5，
Figure BDA0002118756180000141
And isdikRepresenting a vector
Figure BDA0002118756180000142
Modulo or image xiAnd neighbor xjThe euclidean distance between them. The calculation steps of the vector weighted cosine distance method are as follows:
the first step is as follows: solving for gamma, gamma1,Γ2；
The second step is that: solving for
Figure BDA0002118756180000143
And
Figure BDA0002118756180000144
Figure BDA0002118756180000145
the third step: dividing an image x according to an initial composite class parameter xiiTo which new single class or new compound classes consisting of new single classes.
Parameter setting rules: according to the experimental result, N-2 c is recommended as a default value, if N is too small, the performance of CClu is reduced, but if N is too large, many categories are generated, which brings huge calculation amount; xi is in the range of 0,0.4]And is also set to a default value that should be set according to the degree of accuracy that a user may accept during a particular application. At the same time, due to K1And K2Value is [5,10 ]]In between, there is little effect on the experimental results, therefore, K is adjusted1And K2Is set to 5.
In CClu, the true class c is assumed to be either a priori or empirically derived, if no a priori knowledge is available, by
Figure BDA0002118756180000146
Wherein 0 is less than or equal to N (c) is less than or equal to 1.

Claims (9)

1. The image classification method based on the trust function theory is characterized by comprising the following steps of:
dividing all images in an image set X to be classified into a plurality of initial subclasses; the initial sub-class consists of N initial single classes and a plurality of initial composite classes, each initial composite class consists of two initial single classes, N is larger than or equal to c, and c represents the number of real classes of all images in the image set X;
calculating the density value of each initial subclass; wherein the density is the reciprocal of the mean value of Euclidean distances between all images in the initial subclass;
calculating the density of each initial subclass by adopting a formula as follows:
Figure FDA0003166439890000011
where ρ isjIs the jth initial subclass AjThe density of (a) of (b),
Figure FDA0003166439890000012
represents the initial subclass AjImage x of (1)iWith its K in image set X1Mean distance of Euclidean distances of neighbors, K1Is a constant number, njIs an initial subclass AjNumber of middle images, i is the initial subclass AjOrdinal number of the middle image;
combining the images in the initial composite class and the corresponding initial single class according to the magnitude relation of the density values of the initial composite class and the corresponding initial single class to generate a new single class and a new composite class until the sum of the number of the new single class and the initial single class generated after combination is c;
and dividing the images in the initial composite class into the new single class, the initial single class or the new composite class to obtain the classification result of the images in the image set X to be classified.
2. The image classification method based on the trust function theory as claimed in claim 1, wherein the magnitude relationship between the density values of the initial composite class and the corresponding initial single class consists of the following three cases:
the first condition is as follows: rhok,t≥ρkAnd ρk,t≥ρt；
Case two: rhok≤ρk,t≤ρt；
Case three: rhok,t≤ρkAnd ρk,t≤ρt；
Where ρ iskK is the density value of the kth initial single class, k is more than or equal to 0 and less than or equal to N, rhotIs the density value of the t initial single class, t is more than or equal to 0 and less than or equal to N, rhok,tIs the density value of the initial composite class consisting of the kth initial single class and the tth initial single class.
3. The image classification method based on the belief 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 sequentially: case one, case two, case three.
4. The image classification method based on the belief function theory as claimed in claim 3, wherein the specific method of merging the images in the initial composite class and the corresponding initial single class is:
calculating a density distance value between two initial single classes corresponding to the initial composite class, sequencing the initial composite classes from small to large according to the density distance value, and combining the images in the initial composite class and the images in the corresponding initial single classes into a new single class in sequence;
when the density distance values between two initial single classes corresponding to the initial composite classes are equal, sequentially combining the images in the initial composite classes and the images in the corresponding initial single classes into a new single class according to the descending order of the density values of the initial composite classes;
wherein the density distance value is a density set D of two initial single classes in the initial single classsThe initial single-class density set DsIs a set of density values of the N initial single classes, and the initial single class density set DsAccording to the density value of the initial single classAnd sorting by size.
5. The image classification method based on the belief function theory as claimed in claim 4, wherein after a new single class is generated and before the next initial combination of the composite classes, it is determined whether the new single class and other new single classes have the same image, and when the new single class and other new single classes have the same image, all images in the new single class and the new single class having the same image are combined to generate a new single class.
6. The image classification method based on the belief function theory as claimed in any one of claims 2 to 5, wherein the specific method of classifying the images in the initial composite class into the new single class, the initial single class or the new composite class is:
selecting an image x in the initial composite classiAnd finding out K closest to Euclidean distance from the image set X to be classified2The individual image is used as its neighbor image, and K is generated2Vector according to the image xiAnd its adjacent image Euclidean distance between generate K2A weight of the vector;
calculate K2A sum vector of the vectors as a first sum vector;
calculate K2The adjacent image corresponding to the vector end point in the vector belongs to the sum vector of the vectors of the same initial single type/new single type, and the sum vector is used as a second sum vector;
calculating cosine value of included angle formed by each second sum vector and the first sum vector, and displaying the image xiDividing the initial single type/new single type corresponding to the second sum vector with the minimum cosine value of the included angle;
and repeating the steps until the images in all the initial compound classes are divided.
7. The image classification method based on the belief function theory as claimed in claim 6, characterized in that when the difference of cosine values of included angles formed by the two second sum vectors and the first sum vector is smaller than a threshold value, the image is classified 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 belief function theory as claimed in claim 7, characterized in that a c-means clustering algorithm is specifically employed to divide all images in the image set X to be classified into a plurality of initial sub-classes.
9. The image classification method based on the belief function theory as claimed in claim 7 or 8, characterized in that the image set X to be classified is an image set of asymmetric data.
CN201910599618.7A 2019-07-04 2019-07-04 Image classification method based on trust function theory Active CN110472657B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910599618.7A CN110472657B (en) 2019-07-04 2019-07-04 Image classification method based on trust function theory

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910599618.7A CN110472657B (en) 2019-07-04 2019-07-04 Image classification method based on trust function theory

Publications (2)

Publication Number Publication Date
CN110472657A CN110472657A (en) 2019-11-19
CN110472657B true CN110472657B (en) 2021-09-03

Family

ID=68507393

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910599618.7A Active CN110472657B (en) 2019-07-04 2019-07-04 Image classification method based on trust function theory

Country Status (1)

Country Link
CN (1) CN110472657B (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105550714A (en) * 2015-12-30 2016-05-04 国家电网公司 Cluster fusion method for warning information in heterogeneous network environment
CN105930791A (en) * 2016-04-19 2016-09-07 重庆邮电大学 Road traffic sign identification method with multiple-camera integration based on DS evidence theory
CN107368854A (en) * 2017-07-20 2017-11-21 华北电力大学（保定） A kind of circuit breaker failure diagnostic method based on improvement evidence theory

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20080228699A1 (en) * 2007-03-16 2008-09-18 Expanse Networks, Inc. Creation of Attribute Combination Databases

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105550714A (en) * 2015-12-30 2016-05-04 国家电网公司 Cluster fusion method for warning information in heterogeneous network environment
CN105930791A (en) * 2016-04-19 2016-09-07 重庆邮电大学 Road traffic sign identification method with multiple-camera integration based on DS evidence theory
CN107368854A (en) * 2017-07-20 2017-11-21 华北电力大学（保定） A kind of circuit breaker failure diagnostic method based on improvement evidence theory

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
A new pattern classification improvement method with local quality matrix based on K-NN;Zhun-ga Liu et al.;《Knowledge-Based Systems》;20181107;1-27 *
基于置信函数理论的不确定数据分类与决策融合;焦连猛;《中国博士学位论文全文数据库 工程科技Ⅱ辑》;20170815;C028-3 *

Also Published As

Publication number Publication date
CN110472657A (en) 2019-11-19

Similar Documents

Publication Publication Date Title
CN110443281B (en) Text classification self-adaptive oversampling method based on HDBSCAN (high-density binary-coded decimal) clustering
CN107368807B (en) Monitoring video vehicle type classification method based on visual word bag model
Firpi et al. Swarmed feature selection
CN111126482B (en) Remote sensing image automatic classification method based on multi-classifier cascade model
CN107633226B (en) Human body motion tracking feature processing method
CN112633382B (en) Method and system for classifying few sample images based on mutual neighbor
Menardi Density-based Silhouette diagnostics for clustering methods
CN101853389A (en) Detection device and method for multi-class targets
Liu et al. On classification of source cameras: A graph based approach
CN105184260A (en) Image characteristic extraction method, pedestrian detection method and device
CN109150830B (en) Hierarchical intrusion detection method based on support vector machine and probabilistic neural network
Amelio et al. A genetic algorithm for color image segmentation
CN107016416B (en) Data classification prediction method based on neighborhood rough set and PCA fusion
AU2019293020B2 (en) Display control device, display control method, and display control program
CN110472657B (en) Image classification method based on trust function theory
CN106326927B (en) A kind of shoes print new category detection method
CN110929801B (en) Improved Euclid distance KNN classification method and system
CN111027609B (en) Image data weighted classification method and system
CN115511853A (en) Remote sensing ship detection and identification method based on direction variable characteristics
CN109583492A (en) A kind of method and terminal identifying antagonism image
CN114626451A (en) Data preprocessing optimization method based on density
CN112308160A (en) K-means clustering artificial intelligence optimization algorithm
CN111126444A (en) Classifier integration method
CN110647944A (en) Weighted KNN classification method and system based on support vector machine
CN114821206B (en) Multi-modal image fusion classification method and system based on confrontation complementary features

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant