CN110378384B - Image classification method combining privilege information and ordering support vector machine - Google Patents

Image classification method combining privilege information and ordering support vector machine Download PDF

Info

Publication number
CN110378384B
CN110378384B CN201910532935.7A CN201910532935A CN110378384B CN 110378384 B CN110378384 B CN 110378384B CN 201910532935 A CN201910532935 A CN 201910532935A CN 110378384 B CN110378384 B CN 110378384B
Authority
CN
China
Prior art keywords
image
images
negative
positive
prototype
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
CN201910532935.7A
Other languages
Chinese (zh)
Other versions
CN110378384A (en
Inventor
刘倩
刘波
肖燕珊
李松松
刘芷菁
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.)
Guangdong University of Technology
Original Assignee
Guangdong University of Technology
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 Guangdong University of Technology filed Critical Guangdong University of Technology
Priority to CN201910532935.7A priority Critical patent/CN110378384B/en
Publication of CN110378384A publication Critical patent/CN110378384A/en
Application granted granted Critical
Publication of CN110378384B publication Critical patent/CN110378384B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

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
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Image Analysis (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses an image classification method combining privilege information and a sequencing support vector machine, which comprises the steps of firstly constructing a positive marked image set and an unmarked image set, identifying a reliable negative image from the unmarked image set, and utilizing the reliable negative image to establish a positive prototype image and a negative prototype image; deleting reliable negative images from the unlabeled image set, converting the rest unlabeled images, positive images and negative images into vectors respectively, and performing difference on the vectors to form a difference image; representing the difference image by using a similarity model; and constructing a classification model by using the difference image, the similarity weight and the privilege information, and training, wherein the trained classification model is used for classifying the image. In the method, the similarity of the privilege information and the difference image is considered, and the similarity and the privilege information are integrated into a learning stage of the sequencing support vector machine so as to quickly classify the images.

Description

Image classification method combining privilege information and ordering support vector machine
Technical Field
The invention relates to the field of image processing and machine learning, in particular to an image classification method combining privilege information and a sequencing support vector machine.
Background
With the advent of the information age, information data has exploded, and image information expression is vivid and direct, and gradually becomes one of the mainstream information propagation modes. The traditional method of image classification is feature description and detection, and the traditional method may be effective for some simple image classification, but the traditional classification method is not heavy due to the rapid increase of the image data volume, and a quick, efficient and reasonable method is needed for processing and analyzing the images.
The image classification method based on the convolutional neural network becomes a mainstream algorithm of current image classification, and compared with the traditional image classification method, the image classification method does not need to manually describe and extract the characteristics of the target image. The classical convolutional neural network models in common use are of a wide variety, for example, the AlexNet model: the feature description and extraction capability of the image is very limited, and the GoogLeNet model: as the model hierarchy deepens, the gradient dispersion problem becomes more serious, making the network difficult to train, VGGNet model: the convolution kernel used by the convolution layer is smaller, but the depth of the model is increased, and the network uses too many parameters and the training speed is slow. With the development of learning (Positive andunlabeled learning) of the positively labeled examples and unlabeled examples, image classification methods based on PU learning have been proposed, which extend limited training images with unlabeled images to improve the accuracy of the classifier. In PU learning, training examples consist of some positive images and some unlabeled images, which may be positive or negative, with positive and unlabeled images constructing a classifier. Many practical applications today can be seen as PU learning tasks, e.g. information retrieval based on search engine click data, where pages clicked by a user can be as a positive class, while pages not clicked are unlabeled classes.
Existing PU learning methods can be categorized into three categories, namely, two-step strategy-based methods, cost-sensitive methods, and similarity-based methods, depending on the manner in which the unlabeled examples are handled.
However, in many practical applications, it is common for both a positively marked instance and a large number of unmarked instances and their privilege information to be available only during the training phase but not during testing. For example, in image classification, annotations from experts are used as privilege information, such annotations being available only during the training phase but not during testing. In order to solve the problem existing in the three at the same time, a method is needed to effectively use the unlabeled image and also to fully use the privilege information.
Disclosure of Invention
The invention aims to provide an image classification method combining privilege information and a sequencing support vector machine, which considers the existence of privilege information on a similarity PU model and integrates the privilege information into a learning stage of the sequencing support vector machine; the similarity and privilege information of the difference images are considered and are integrated into the learning stage of the sequencing support vector machine so as to quickly classify the images.
In order to realize the tasks, the invention adopts the following technical scheme:
an image classification method combining privilege information and a sequencing support vector machine comprises the following steps:
step 1, randomly selecting images from an image data set to construct positive and negative classes, randomly selecting part of the images from all the images in the positive class to mark, marking the part of the images as marked positive images, and forming a positive mark image set; forming an unlabeled image set from the rest images in the positive class and all the images in the negative class;
step 2, identifying reliable negative images from the unlabeled images, and establishing a positive prototype image and a negative prototype image by utilizing the reliable negative images;
step 3, deleting the reliable negative image from the unlabeled image set, then respectively converting the rest unlabeled image, positive image and negative image into vectors, and making difference between the vectors to form a difference image;
representing the difference image by using a similarity model;
and 4, constructing a classification model by using the difference image, the similarity weight and the privilege information, and training, wherein the trained classification model is used for classifying the image.
Further, the identifying reliable negative images from the unlabeled image set includes:
and respectively utilizing the Spy technology and the Rocchio technology to identify the reliable negative images from the unlabeled image set, taking the reliable images primarily identified by the Spy technology and the Rocchio technology as intersections, and taking the images in the intersections as final reliable negative images.
Further, the creating representative positive and negative prototype images with reliable negative images includes:
the method comprises the steps of setting a reliable negative images, dividing the reliable negative images into m micro-clusters by using a k-means clustering method to establish a positive prototype image and a negative prototype image, wherein the rest unlabeled images are u', and the calculation formula is as follows:
positive prototype image:
negative prototype image:
wherein, the parameters alpha, beta and t are adjusting parameters,p v representing a positive prototype image representative of the v-th micro-cluster, n v Representing a negative prototype image representative of the v-th micro-cluster; PS denotes a positive marked image set, US denotes an unmarked image set, NS v Representing a reliable negative image set of the v-th micro-cluster.
Further, the difference image is expressed as:
X e ＝x i -x j ,X q ＝x i -x k ,X n ＝x k -x j ；
wherein x is i ∈PS,x j ∈NS,x k ∈US'，X e ,X q ,X n Is the difference image formed, US' represents the set of images remaining after the reliable negative image is deleted from the unlabeled image set.
Further, the method for representing the difference image by using the similarity model is specifically expressed as follows:
{X,(m + (X),m - (X))}
wherein X represents a difference image, m + (X) and m - (X) is a similarity weight representing the similarity of the difference image X to the positive and negative classes, respectively; d, d 1 、d 2 The Euclidean distance from the difference image to the representative positive prototype image and the representative negative prototype image; the selection method of the representative positive prototype image and the representative negative prototype image comprises the following steps:
calculating Euclidean distance from each positive prototype image to each negative prototype image, and selecting d for maximizing Euclidean distance max A pair of positive and negative prototype images serves as a representative positive prototype image and a representative negative prototype image.
Further, the classification model is expressed as:
in the above formula:
X e ＝x i -x j ,X q ＝x i -x k ,X n ＝x k -x j ；S 1 ＝p·a,S 2 ＝p·u',S 3 ＝u'·a，/>where a represents the number of reliable negative images extracted, u' represents the number of unlabeled images remaining, p represents the number of labeled positive images, x i ∈PS,x j ∈NS,x k ∈US'，/>Privilege information indicating positive image, negative image, remaining unlabeled image, respectively, ++>Privilege information, w and w representing difference images * The weight vector is represented by a weight vector,<,>represent the inner product, C 1 、C 2 And C 3 Is a penalty factor controlling the trade-off between hyperplane edges and errors, is an error term, gamma is a non-negative parameter, parameter θ e ，θ q And theta n Control error items +.>Is a coefficient of (a).
The invention has the following technical characteristics:
1. the scheme refers to the fact that the prediction is also positive, and the prediction is called True (TP); the actual positive class, predicted negative class, called False Negative (FN); the actual negative class, predicted positive class, called False Positive (FP); the actual negative class, the predicted negative class, is also called True Negative (TN). In the analysis of anomaly classification models, the main concern is that anomalies (i.e., true negative rates) are correctly detected,the classification accuracy of the classifier is improved by using unlabeled images and similarity weights, so that the number of instances of TNs is increased, the number of instances of FPs is reduced, the true negative rate is slightly increased, namely the probability of correctly detecting abnormality is increased, and the system can be timely repaired. It is naturally important to reduce the number of instances that are mispredicted as abnormal instances (false negative rates), which in fact have little impact on the system. In a true negative rate sample, as opposed to a false negative rate, information about the correct identification of the abnormal instance is contained.
2. The invention is proposed in the case of the simultaneous existence of PU problem and privilege information. Firstly, learning by considering the added privilege information and the common training image, and defining a correction function and a relaxation variable in a privilege feature space to evaluate the relaxation variable in the standard support vector machine. And then the ranking information is integrated into a learning stage of a ranking support vector machine, and the ranking support vector machine is used for effectively training ranking, so that compared with a standard support vector machine, the ranking information can solve the same optimization problem and is faster. Since a rank support vector machine is used, the similarity weights are for the difference images.
Drawings
FIG. 1 is a schematic flow chart of the method of the present invention.
Detailed Description
When the method is used in the image classification and identification process, positive and negative classes are constructed according to the image data set; for all images (instances) in the positive class, a small portion of them is randomly selected as marked positive images, while the other images in the positive class and all images in the negative class constitute an unmarked image set. Thus, a set of positively marked and unmarked images can be obtained:p represents the number of positive mark images, x i Representing a positive mark image; />u represents the number of unlabeled images, x k Representing an unlabeled image. The marked positive images belong to the class of high correlation, while the unmarked images belong to the class of lower correlation. All images are adjusted to 10 multiplied by 10 pixels, each image takes 100-dimensional original pixel vectors as main feature vectors, each training image is added with a description, and the training image is converted into 21-dimensional text feature vectors and is used as privilege information; the goal of this scheme is to construct a classifier using positively labeled and unlabeled images, and then classify the images using the classifier.
An image classification method combining privilege information and a sequencing support vector machine has the basic idea that the most reliable negative image NS is extracted from unlabeled examples, and then NS is deleted from the unlabeled images, namely US' =US-NS. Then image aggregation in NS using K-means clusteringClass m micro-clusters, denoted NS 1 ，NS 2 ，L,NS m To build a representative positive prototype and a representative negative prototype; generating two similarity weights m for difference images + (x) And m - (x) The method comprises the steps of carrying out a first treatment on the surface of the The difference examples formed by the positive image set, the negative image set and the unlabeled image set, and similar weight and privilege information are integrated into a learning model of a sequencing support vector machine. The specific steps of the invention are as follows:
step 1, constructing a positive marked image set and an unmarked image set
Randomly selecting some images from the image dataset to construct a positive class and a negative class, wherein the images in the positive class are called positive images, and the images in the negative class are called negative images; for all images in the positive class, randomly selecting part of the images for marking, marking the images as marked positive images, and forming a positive marking image set; the remaining images in the positive class and all images in the negative class are formed into an unlabeled image set.
For example, in the MNIST dataset, the scheme selects a digital image of "0" as the positive class, and the remaining digital images as the negative classes. For all positive images in the positive class, the scheme randomly selects 30% of them as marked positive images, i.e. marks the 30% of positive class images as +1, while the other images in the positive class and all negative images in the negative class constitute an unlabeled image set. Thus, a positive marked image set and an unmarked image set can be obtained.
Step 2, establishing representative positive prototype and negative prototype images
Using the Spy technique (Spy technology, [ Liu et al, 2002.)]) And Rocchio technology (Rocchio technique, [ Li and Liu, 2003)]) Identifying a reliable negative image from the unlabeled image set, denoted N 1 ，N 2 The method comprises the steps of carrying out a first treatment on the surface of the Only if both techniques consider them as negative images, they are considered as reliable negative images, namely: ns=n 1 ⌒N 2 。
The reliable negative images obtained in the unlabeled image set are provided with a pieces, the rest unlabeled images are u pieces, the a reliable negative images are divided into m micro-clusters by a k-means clustering method, so that a positive prototype image and a negative prototype image are established, and the calculation formula is as follows:
positive prototype image:
negative prototype image:
wherein x represents norms of the image x; the adjustment parameters alpha and beta are set to 16 and 4,t is an adjustment parameter, generally set to 30; p is p v Representing a positive prototype image representative of the v-th micro-cluster, n v Representing a negative prototype image representative of the v-th micro-cluster; PS denotes a positive marked image set, US denotes an unmarked image set, NS v Representing a reliable negative image set of the v-th micro-cluster.
Step 3, constructing a difference image and calculating the similarity weight of the difference image
Deleting reliable negative images from the unlabeled image set, and marking the set of the rest unlabeled images as US '(US' =us-NS); converting the positive image and the negative image into vectors by using an HOG feature extraction method, and marking the left unlabeled images as positive labeled vectors, negative class vectors and unlabeled class vectors;
generating a difference image represented by the difference vector by differencing the positive label vector, the negative class vector and the rest of the unlabeled class vectors, namely, differencing the positive label vector and the negative class vector, differencing the positive label vector and the unlabeled class vector, differencing the unlabeled vector and the negative class vector, such as:
X e ＝x i -x j ,X q ＝x i -x k ,X n ＝x k -x j ；
wherein x is i ∈PS,x j ∈NS,x k ∈US'，X e ,X q ,X n Is the difference image formed. Since it is not known which class the difference image belongs to, the scheme introduces a similarity model { X, (m) + (X),m - (X)) } to represent a difference image X (X is represented by X) e 、X q Or X n ). In order to calculate the similarity weight of the difference image, the scheme provides a similarity weight generation method based on a distance similarity mechanism:
the scheme firstly calculates each positive prototype image p v To each negative prototype image n v Is selected such that the Euclidean distance is a maximum d max Is denoted as p max ,n max The method comprises the steps of carrying out a first treatment on the surface of the Respectively calculating difference images to representative positive prototype images p max Representative negative prototype image n max Euclidean distance d of (2) 1 、d 2 The method comprises the steps of carrying out a first treatment on the surface of the The corresponding weights for the difference image are calculated as follows:
wherein m is + (X) and m - (X) is a similarity weight which respectively represents the similarity of the difference image X with the positive class and the negative class, and m is more than or equal to 0 + (X) is less than or equal to 1 and 0 less than or equal to m - (X) is less than or equal to 1. If the difference image is of the positive class, then m + (X) is equal to 1, corresponding m - (X) equals 0 and vice versa, m if negative + (X) is equal to 0, m - (X) is equal to 1.
And 4, constructing a classification model by utilizing the difference image, the similar weight and the privilege information, and realizing the learning and classification of the image:
the scheme combines the difference image, the similarity weight and the privilege information thereof into a learning stage of a sequencing support vector machine so as to construct a more accurate classifier.
Is provided withIs a new training set of images formed from the difference images, their labels and privilege information. Training equation 1 on the new image training set, and obtaining the optimal ranking function. Wherein X is l Representing a difference image, and X l Related additional information->Is privilege information, for example, each training image in MNIST data set is added with an overall description, and an expert converts the training image into a 21-dimensional feature vector and takes the feature vector as privilege information; z l X represents l S (s=p·u) represents the number of images in the training set, where p represents the number of marked positive images in the image dataset and u represents the number of unmarked images in the original image dataset. Assuming that both the ordering function and the correction function are linear, i.e. f (X) =<w,X>＝w T X (T represents a transpose) and phi (X * )＝<w * ,X * >Wherein w, w * The weight is represented by a weight that,<,>representing the inner product. We find the best ranking function from a set of allowed functions, which ensures that the error classification rate is the lowest. The problem of requiring optimization can therefore be expressed as:
in formula 1:
X e ＝x i -x j ,X q ＝x i -x k ,X n ＝x k -x j ；S 1 ＝p·a,S 2 ＝p·u',S 3 ＝u'·a，/>where a represents the number of reliable negative images extracted, u' represents the number of unlabeled images remaining, p represents the number of labeled positive images, x i ∈PS,x j ∈NS,x k ∈US'，Privilege messages respectively representing positive image, negative image and remaining unlabeled imageRest, I/II>Privilege information (superscript x indicates privilege information) indicating a difference image. w and w * The weight vector is represented by a weight vector,<w,w>and<w * ,w * >privilege information X for dividing difference image X and difference image * Are the regular terms of (2) to prevent overfitting;<,>represent the inner product, C 1 、C 2 And C 3 Is a penalty factor controlling the trade-off between hyperplane edge and error,/->Is an error term of the error term,is a correction function, f (X e )＝<w,X e >,f(X q )＝<w,X q >,f(X n )＝<w,X n >Respectively difference image X e ,X q ,X n Ordering function of->And->Errors that can be considered to be of different weights; gamma is a non-negative parameter reflecting the proportion of privilege information in the classifier model, parameter θ e ，θ q And theta n Control error items +.> Is a coefficient of (a).
By combining Lagrangian functions with the original variables w, w * 、Differentiation is performed to obtain the formula1 to dual form 2:
wherein alpha is e ≥0，α q ≥0，α n ≥0，β e ≥0，β q ≥0，β n ≥0，λ e ≥0，λ q Not less than 0 and lambda n And ≡0 is Lagrangian multiplier. Parameter θ e ，θ q And theta n Respectively control error termsIs a coefficient of (a).
Solving parameters w, w * ：
The Lagrangian multiplier alpha can be calculated for the formula 2 by utilizing the SMO algorithm e ,α q ,α n ,β e ,β q ,β n W and w can be obtained by using the following formula * Is the optimal solution of (a):
and after the optimal solution is obtained, an optimal sorting function (optimal classifier) can be obtained, and the images in the image data set can be sorted by utilizing the optimal sorting function, so that sorting is performed. The sorting function value of the positive class image is higher than that of the negative class image, namely the correlation of the positive class image is higher than that of the negative class image.
For MNIST datasets, all images are first turned into feature vectors by existing HOG feature extraction methods. Centralizing image dataThe vectors are subtracted from each other to form a disparity vector (representing a disparity image X bc ＝x b -x c ). Then taking the difference image as the input of the optimal sorting function, wherein the prediction label of the difference image is as follows:
if w T X bc Not less than 0, x b ＞ x c I.e. image x b Is higher than image x c Is a correlation of (3). According to the correlation of the positive class image being higher than that of the negative class image, x can be determined b Is a positive image, x c Is a negative image. If the number 0 is considered as positive in the experiment, then image x b For the image number 0, if the number 1 is considered as positive in the experiment, then the image x b Image number 1, and so on. I.e. classification is completed.
For PU learning problems, the method is different from a two-step strategy classification method, unlabeled images are excluded in a learning stage, and the classifier is trained by using positive and negative images only, so that many researches show that the unlabeled images are more easily positioned near decision boundaries and play a critical role in the construction of the classifier, and the method reduces the performance of the classifier. The method considers the condition that the positive marked image and a large amount of unmarked images and privilege information thereof exist simultaneously, and in addition, the positive marked images and the unmarked images are not integrated into the learning phase of the standard support vector machine, but are sequenced into the learning phase of the support vector machine. For PI learning problems, the method presented herein differs from the classical paradigm of LUPI in that rather than integrating the training image set into the learning phase of a standard support vector machine, it orders the support vector machine learning phase.

Claims (5)

1. An image classification method combining privilege information and a sequencing support vector machine is characterized by comprising the following steps:
step 1, randomly selecting images from an image data set to construct positive and negative classes, randomly selecting part of the images from all the images in the positive class to mark, marking the part of the images as marked positive images, and forming a positive mark image set; forming an unlabeled image set from the rest images in the positive class and all the images in the negative class;
step 2, identifying reliable negative images from the unlabeled images, and establishing a positive prototype image and a negative prototype image by utilizing the reliable negative images;
step 3, deleting the reliable negative image from the unlabeled image set, then respectively converting the rest unlabeled image, positive image and negative image into vectors, and making difference between the vectors to form a difference image;
representing the difference image by using a similarity model;
step 4, constructing a classification model by using the difference image, the similarity weight and the privilege information, and training, wherein the trained classification model is used for classifying the image;
is provided withIs a new image training set formed by the difference image, the label and the privilege information thereof;
training 1 on the new image training set, and solving an optimal sorting function; wherein X is l Representing a difference image, and X l Related additional informationIs privilege information; z l X represents l (s=p·u represents the number of images in the training set, where p represents the number of marked positive images in the image dataset and u represents the number of unmarked images in the original image dataset;
assuming that both the ordering function and the correction function are linear, i.e. f (X) =<w,X>＝w T X, T represents a transpose; phi (X) * )＝<w * ,X * >Wherein w, w * The weight is represented by a weight that,<,>representing the inner product; the problem of optimization to be solved is expressed as:
in formula 1:
S 1 ＝p·a，S 2 ＝p·u'，S 3 ＝u'·a，where a represents the number of reliable negative images extracted, u' represents the number of unlabeled images remaining, p represents the number of labeled positive images, x i ∈PS,x j ∈NS,x k ∈US'，/>Privilege information respectively representing a positive image, a negative image, and the remaining unlabeled image,privilege information indicating a difference image, and superscript indicates privilege information; w and w * The weight vector is represented by a weight vector,<w,w>and<w * ,w * >privilege information X of difference image X and difference image X, respectively * Are the regular terms of (2) to prevent overfitting;<,>represent the inner product, C 1 、C 2 And C 3 Is a penalty factor controlling the trade-off between hyperplane edges and errors,is an error term->Is a correction function, f (X e )＝<w,X e >,f(X q )＝<w,X q >,f(X n )＝<w,X n >Respectively difference image X e ,X q ,X n Is used as a ranking function of (a),and->Considered as error of different weights, X represents difference image, m + (X) and m - (X) is a similarity weight representing the similarity of the difference image X to the positive and negative classes, respectively; gamma is a non-negative parameter reflecting the proportion of privilege information in the classifier model, parameter θ e ，θ q And theta n Control error items +.>Coefficients of (2);
by combining Lagrangian functions with the original variables w, w * 、Differentiation is performed to convert equation 1 to dual form 2:
s.t.
C 1 θ e -α e -λ e ＝0
C 2 m - (X q )θ q -α q -λ q ＝0
C 3 m - (X n )θ n -α n -λ n ＝0
0≤α e ≤θ e C 1 ,0≤α q ≤θ q C 2 ,0≤α n ≤θ n C 1
… … type 2
Wherein alpha is e ≥0，α q ≥0，α n ≥0，β e ≥0，β q ≥0，β n ≥0，λ e ≥0，λ q Not less than 0 and lambda n 0 is Lagrangian multiplier; parameter θ e ，θ q And theta n Respectively control error termsCoefficients of (2);
solving parameters w, w * ；
The Lagrangian multiplier alpha can be calculated for the formula 2 by utilizing the SMO algorithm e ,α q ,α n ,β e ,β q ,β n W and w can be obtained by using the following formula * Is the optimal solution of (a):
and obtaining an optimal sorting function after obtaining the optimal solution, and sorting the images in the image data set by using the optimal sorting function so as to sort the images.
2. The method for image classification in combination with a privilege information and ordering support vector machine of claim 1 wherein said identifying reliable negative images from a set of unlabeled images comprises:
and respectively utilizing the Spy technology and the Rocchio technology to identify the reliable negative images from the unlabeled image set, taking the reliable images primarily identified by the Spy technology and the Rocchio technology as intersections, and taking the images in the intersections as final reliable negative images.
3. The method of claim 1, wherein said creating representative positive and negative prototype images from reliable negative images comprises:
the method comprises the steps of setting a reliable negative images, dividing the reliable negative images into m micro-clusters by using a k-means clustering method to establish a positive prototype image and a negative prototype image, wherein the rest unlabeled images are u', and the calculation formula is as follows:
positive prototype image:
negative prototype image:
wherein, the parameters alpha, beta and t are adjusting parameters,p v representing a positive prototype image representative of the v-th micro-cluster, n v Representing a negative prototype image representative of the v-th micro-cluster; PS denotes a positive marked image set, US denotes an unmarked image set, NS v Representing a reliable negative image set of the v-th micro-cluster.
4. The method for classifying images in combination with privilege information and ordered support vector machine as claimed in claim 1, wherein said difference image is expressed as:
X e ＝x i -x j ,X q ＝x i -x k ,X n ＝x k -x j ；
wherein x is i ∈PS,x j ∈NS,x k ∈US'，X e ,X q ,X n Is the difference image formed, US' represents the set of images remaining after the reliable negative image is deleted from the unlabeled image set.
5. The method for classifying images in combination with privilege information and a sequencing support vector machine as claimed in claim 1, wherein said representing said difference image by using a similarity model is specifically:
{X,(m + (X),m - (X))}
wherein X represents a difference image, m + (X) and m - (X) is a similarity weight representing the similarity of the difference image X to the positive and negative classes, respectively; d, d 1 、d 2 The Euclidean distance from the difference image to the representative positive prototype image and the representative negative prototype image; the selection method of the representative positive prototype image and the representative negative prototype image comprises the following steps:
calculating Euclidean distance from each positive prototype image to each negative prototype image, and selecting d for maximizing Euclidean distance max A pair of positive and negative prototype images serves as a representative positive prototype image and a representative negative prototype image.
CN201910532935.7A 2019-06-19 2019-06-19 Image classification method combining privilege information and ordering support vector machine Active CN110378384B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910532935.7A CN110378384B (en) 2019-06-19 2019-06-19 Image classification method combining privilege information and ordering support vector machine

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910532935.7A CN110378384B (en) 2019-06-19 2019-06-19 Image classification method combining privilege information and ordering support vector machine

Publications (2)

Publication Number Publication Date
CN110378384A CN110378384A (en) 2019-10-25
CN110378384B true CN110378384B (en) 2023-10-10

Family

ID=68249330

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910532935.7A Active CN110378384B (en) 2019-06-19 2019-06-19 Image classification method combining privilege information and ordering support vector machine

Country Status (1)

Country Link
CN (1) CN110378384B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
TWI743837B (en) * 2020-06-16 2021-10-21 緯創資通股份有限公司 Training data increment method, electronic apparatus and computer-readable medium
CN111967500B (en) * 2020-07-21 2023-10-13 广东工业大学 Classification method combining non-parallel dual support vector machine and sample privilege information

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102750347A (en) * 2012-06-08 2012-10-24 天津大学 Method for reordering image or video search
CN109165673A (en) * 2018-07-18 2019-01-08 广东工业大学 Image classification method based on metric learning and more example support vector machines
CN109299162A (en) * 2018-11-08 2019-02-01 南京航空航天大学 A kind of Active Learning Method classified for positive class and data untagged

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102750347A (en) * 2012-06-08 2012-10-24 天津大学 Method for reordering image or video search
CN109165673A (en) * 2018-07-18 2019-01-08 广东工业大学 Image classification method based on metric learning and more example support vector machines
CN109299162A (en) * 2018-11-08 2019-02-01 南京航空航天大学 A kind of Active Learning Method classified for positive class and data untagged

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
Learning to Rank Using Privileged Information;Viktoriia Sharmanska，等;《IEEE International Conference on Computer Vision》;20131231;论文第825页第15栏第1行至第832页第1栏第19行 *
Similarity-Based Approach for Positive and Unlabelled Learning;Yanshan Xiao，等;《International Joint Conference on Artificial Intelligence》;20111231;论文第1577页第1栏第29行至第1581页第1页第34行 *
基于正样本和未标记样本的遥感图像分类方法;裔阳，等;《计算机工程与应用》;20170228;第54卷(第4期);论文第160页第1栏倒数第3行至第166页第1栏第5行 *

Also Published As

Publication number Publication date
CN110378384A (en) 2019-10-25

Similar Documents

Publication Publication Date Title
Liu et al. Connectionist temporal classification with maximum entropy regularization
US7885466B2 (en) Bags of visual context-dependent words for generic visual categorization
Eigen et al. Nonparametric image parsing using adaptive neighbor sets
CN111126488B (en) Dual-attention-based image recognition method
CN107683469A (en) A kind of product classification method and device based on deep learning
CN105760888B (en) A kind of neighborhood rough set integrated learning approach based on hierarchical cluster attribute
CN103425996B (en) A kind of large-scale image recognition methods of parallel distributed
Tarawneh et al. Invoice classification using deep features and machine learning techniques
JP2003256441A (en) Document classification method and apparatus
CN110647907B (en) Multi-label image classification algorithm using multi-layer classification and dictionary learning
CN108427740B (en) Image emotion classification and retrieval algorithm based on depth metric learning
JP6897749B2 (en) Learning methods, learning systems, and learning programs
Tasci Voting combinations-based ensemble of fine-tuned convolutional neural networks for food image recognition
CN110008365B (en) Image processing method, device and equipment and readable storage medium
Freytag et al. Labeling examples that matter: Relevance-based active learning with gaussian processes
CN108446334A (en) Image retrieval method based on content for unsupervised countermeasure training
Schinas et al. CERTH@ MediaEval 2012 Social Event Detection Task.
CN104091038A (en) Method for weighting multiple example studying features based on master space classifying criterion
CN110378384B (en) Image classification method combining privilege information and ordering support vector machine
CN109299263A (en) File classification method, electronic equipment and computer program product
CN111062510A (en) Civil aviation engine blade crack prediction method based on transfer learning
CN116525075A (en) Thyroid nodule computer-aided diagnosis method and system based on few sample learning
CN113076490B (en) Case-related microblog object-level emotion classification method based on mixed node graph
Kwolek et al. Breast cancer classification on histopathological images affected by data imbalance using active learning and deep convolutional neural network
Liu et al. Few-Shot Object Detection in Remote Sensing Images via Label-Consistent Classifier and Gradual Regression

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