CN114580566A - Small sample image classification method based on interval supervision contrast loss - Google Patents

Small sample image classification method based on interval supervision contrast loss Download PDF

Info

Publication number
CN114580566A
CN114580566A CN202210289086.9A CN202210289086A CN114580566A CN 114580566 A CN114580566 A CN 114580566A CN 202210289086 A CN202210289086 A CN 202210289086A CN 114580566 A CN114580566 A CN 114580566A
Authority
CN
China
Prior art keywords
sample
samples
contrast loss
image
support
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.)
Withdrawn
Application number
CN202210289086.9A
Other languages
Chinese (zh)
Inventor
杨赛
胡彬
杨慧
周伯俊
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.)
Nantong University
Original Assignee
Nantong 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 Nantong University filed Critical Nantong University
Priority to CN202210289086.9A priority Critical patent/CN114580566A/en
Publication of CN114580566A publication Critical patent/CN114580566A/en
Withdrawn legal-status Critical Current

Links

Images

Classifications

    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to the technical field of small sample image classification, in particular to a small sample image classification method based on interval supervision contrast loss, which comprises the following steps: pre-training a model on a base class data set by using a novel interval supervision contrast loss function, fixing parameters in an encoder in the pre-training model, extracting characteristics of support image samples in the new class data set, training an SVM classifier, and finally performing classification decision on query samples by using an SVM. The interval supervision contrast loss function in the invention establishes a mathematical model for the contrast relation between the base class samples, rather than only paying attention to the class to which the base class samples belong, and the pre-trained backbone network has more mobility. The supervision contrast loss function in the invention can further reduce the distance of data in the class and increase the distance between classes by increasing the interval parameters, thereby further improving the classification performance.

Description

Small sample image classification method based on interval supervision contrast loss
Technical Field
The invention relates to the technical field of small sample image classification, in particular to a small sample image classification method based on interval supervision contrast loss.
Background
In recent years, deep learning has made a significant breakthrough in several fields of artificial intelligence, such as computer vision, speech recognition, natural language processing, automated driving, etc. However, the powerful learning capabilities of deep convolutional neural networks rely entirely on large amounts of manual label data. The large data-driven nature of the technique severely limits the application of deep learning techniques in many practical situations. Because the cost of accurately labeling large amounts of data is very costly and sometimes impossible, such as medical data for rare diseases, image data for rare species, etc. Under such a background, researchers have actively studied the problem of classifying small sample images, and the purpose of the research is to complete classification and recognition of images by a deep neural network when the number of samples per category is very small.
Meta learning obtains intrinsic meta knowledge by learning a large number of tasks, so that new tasks can be rapidly processed, and the meta learning is a mainstream learning paradigm for solving the problem of small sample image classification. Recent studies, however, have shown that combining feature extractors pre-trained on the entire base class dataset with traditional classifiers can achieve classification performance comparable to complex meta-learning models. The method adopts a two-stage transfer learning paradigm, namely, a backbone network is trained on a large number of labeled base class data sets, so that a good feature extractor can be improved for new class data with a small number of samples in the next class post-processing stage. Generally, the method uses a cross entropy loss function to perform optimization training, and then removes the last layer related to the base class in the backbone network, extracts new class features and completes classification decision. For example, Chen et al (ChenWY, Liu C Y, Kira Z, Wang Y F, Huang J B.A custom look at raw-shot classification [ C ]// Proceedings of the 7th International Conference on Learning retrieval. New Oreans, USA: ICLR,2019:1-8.) propose to optimize the neural network consisting of backbone network and Cosine classifier at the base class dataset using cross entropy loss function, then fix backbone network parameters and fine tune the Cosine classifier with the new class support samples to complete the classification of the query samples. Subsequently, several efforts were proposed in succession, how to further improve the generalization performance of backbone networks. For example, Liu et al (Liu B, Cao Y, Lin Y T, Li Q, Zhang Z, Long M S, Hu H. Positive mapping tables: understating mapping in raw-shot classification [ C ]// Proceedings of the16th European Conference on Computer vision. Berlin, German: Springer,2020:1-8.) also employed a model consisting of backbone network and Cosine classifier in the pre-training phase, while optimizing using a cross-entropy loss function with spacing; mangla et al (Mangla P, Kumari N, Sinha A, Singh M, Krishn acupressure B, Balasuramanian V N. charting the right maniffold: Maniffold mix for feed-shot sharing [ C ]// Proceedings of the IEEE Winter Conference on Applications of Computer Vision, Colorado, USS,2020: 2218-2227.) and Tian et al (Tian Y, Wang Y, Krispann D, et al. Retening few-shot image classification: a good matching you neu/Proceedings of the16th year Conference on company, jet in which distillation was proposed in the future: spring et al. prediction technology: spring, in which A. was introduced in the prediction process and the self-supervision process 2020, respectively; rizve et al (Rizve M N, Khan S, Khan F S, expanding computational Strengths of Invariant and equivalent responses for Few-Shot Learning [ C ]// Proceedings of the 33rd IEEE Conference on Computer Vision and Pattern Recognition.) design a multitask Learning loss function that can maintain isotopy and invariance to complete the pre-training of backbone networks.
After the pre-training phase is completed, only the backbone network in the model is retained since the last layer associated with the base class is removed. At this time, if the classification and identification of the new class data are to be completed, the decision is completed by means of the retrained classifier. The nonparametric model K-nearest neighbor classifier or prototype classifier is a common classification model in a small sample image classification task due to simplicity and effectiveness. For example, Wang et al (Wang Y, Chao W L, Weinberger K Q, Maaten V D. Simpleshot: revising near-neighbor classification for raw-shot learning. arXivpreprint arXiv:1911.04623,2019:1-8.) can achieve good classification performance using a KNN classifier after whitening preprocessing of new-type data. Wang et al (Wangz Y, ZhaoYF, Li J, TianYH. cooperative bi-path metric for raw-shot learning [ C ]// Proceedings of the 28th ACM International Conference on multimedia. New York, USA: ACM press,2020: 1524. sup. 1532.) also used KNN classifiers, but used query samples and correlations between support and base classes to determine neighbor samples in the classifier. Besides, parameter classifiers such as SVM and logistic regression also achieve good classification performance in a small sample image classification task.
The above method for classifying the thumbnail images by adopting the transfer learning paradigm achieves classification performance comparable to that of a complex meta-learning model. Generally, such methods use a cross entropy loss function to pre-train the whole base class data set, and pay more attention to the class to which the base class sample belongs, thereby seriously affecting the mobility of the base class sample in a new class image domain.
In order to solve the above problems, the present invention provides a small sample image classification method based on interval supervision contrast loss.
Disclosure of Invention
Aiming at the problems, the invention provides a small sample image classification method based on interval supervision and contrast loss, which leads a backbone network to have mobility by introducing a supervision and contrast loss function with intervals to pre-train on a base class data set, thereby improving the accuracy of identifying a new class of data samples.
In order to achieve the purpose, the technical scheme adopted by the invention is as follows:
a small sample image classification method based on interval supervision contrast loss comprises the following steps:
step 1: for a given dataset D, determining a base class dataset D to be processedbaseAnd a new class data set Dnovel(ii) a Wherein
Figure BDA0003559427200000021
xiRepresents the ith base class image sample, y iIndicates its corresponding label; at DnovelThe N-way-K-shot classification tasks are constructed, and each classification task is provided with a support sample set
Figure BDA0003559427200000022
And querying the sample set
Figure BDA0003559427200000023
Composition of (a) wherein xsDenotes the s-th support sample, ysDenotes the corresponding label, xqRepresents the qth query sample;
step 2: from base class dataset DbaseRandomly decimating M image samples, expanding them to 2M samples using image enhancement techniques, and inputting them to an encoder E with a parameter ααProjector P with parameter betaβIn (c), a set of 2M data features is obtained
Figure BDA0003559427200000031
Assuming any data in the set as an anchor point and taking the anchor point as a positive sample, taking all samples in the set A (i) with the anchor point label as positive samples to construct a positive sample set P (i), and taking all the rest samples in the set A (i) as negative samples to construct a negative sample set N (i);
and step 3: calculating the interval supervision contrast loss between the positive sample set P (i) and the negative sample set N (i), calculating the gradient of the loss function relative to the parameter, and using gradient descent algorithm to the encoder EαMiddle parameter alpha and projector PβOptimizing the parameter beta in the (-) step;
step 4, pre-training the coder E in the step 3 α(. the) is used as a feature extractor to extract features of the support image samples in the new class data set and train the SVM;
and 5: pre-trained encoder E in step 3αAnd the (-) is used as a feature extractor for extracting features of the query image samples in the new class data set, and the trained SVM is used for testing.
Preferably, the specific steps of step 1 are as follows:
s11: for a given data set D, it is partitioned into three subsets, the training set DtrainVerification set Dval and test set DtestThe classification categories in each subset are not the same, namely:
Figure BDA0003559427200000032
Dtrain∪Dval∪Dtest＝D；
s12: will train set DtrainAs base class dataset DbaseTest set Dtest as a new class data set Dnovel(ii) a At DnovelRandomly extracting N classes in the data set, and randomly extracting K samples in each class to obtain a support sample set
Figure BDA0003559427200000033
Extracting a batch of samples from the residual data in the N categories to obtain a query sample set
Figure BDA0003559427200000034
Preferably, the specific steps of step 2 are as follows:
s21: from DtrainExtracting M image samples as a training batch image sample set DMFor the ith base class image sample x thereiniEnhanced by enhancement techniques
Figure BDA0003559427200000035
And adding it to the batch image sample set;
s22: data set DMIs input to encoder E α(. and projector P)β(. o) obtain the corresponding d-dimensional feature vector, where encoder Eα(. to) ResNet12, projector Pβ() is a multi-layer sensor with only one hidden layer;
then the ith base class image sample xiIs characterized by:
ui＝Pβ(Eα(xi))；
s23: through the feature extraction link, a set consisting of 2M data features is obtained at the moment
Figure BDA0003559427200000036
Will uiAs a positive example, the labels and y of all samples in set A (i)iAll samples are regarded as positive samples to form a positive sample set P (i), and the rest samples are regarded as negative samples to form negative samplesThis set N (i).
Preferably, the specific steps of step 3 are as follows:
s31: the calculation formula of the interval supervision contrast loss function constructed based on the positive example sample set P (i) and the negative example sample set N (i) is as follows:
Figure BDA0003559427200000041
wherein the content of the first and second substances,
Figure BDA0003559427200000042
in order to be a temperature coefficient of the temperature,
Figure BDA0003559427200000043
is a spacing coefficient, upIs the p-th sample in the positive sample set P (i), uaFor negative sample set N (i), s (-) represents the cosine similarity measure of any two samples, and the expression is:
Figure BDA0003559427200000044
s32: calculating the gradient of the interval supervision contrast loss relative to the parameter, wherein the calculation formula is as follows:
Figure BDA0003559427200000045
wherein
Figure BDA0003559427200000046
Figure BDA0003559427200000047
Figure BDA0003559427200000048
S33: and updating the parameters by using a gradient descent algorithm until the algorithm converges, and finishing the pre-training of the backbone network at the moment.
Preferably, the specific steps of step 4 are as follows:
s41: fixed encoder EαThe parameter alpha in the (-) extracts the feature of the support sample in the new class data set, and the d-dimensional feature of the s-th support sample is expressed as:
us＝Pβ(Eα(xs))
s42: selecting a linear support vector machine to classify and judge the new data samples, wherein an optimized objective function is as follows:
Figure BDA0003559427200000049
s.t.ys(w·us+b)≥1-ξs
ξs≥0,s＝1,2,…,2K
wherein, ysIs a label corresponding to the s-th support sample, C is a penalty coefficient, b is a classification threshold value, and xisAnd (4) optimizing and solving the above formula to obtain L support vectors, wherein the w is a parameter of the classifier and is a relaxation variable.
Preferably, the specific steps of step 5 are as follows:
s51: fixed encoder EαThe parameter alpha in the (-) extracts the feature of the query sample in the new class data set, and the d-dimensional feature of the qth support sample is expressed as:
uq＝Pβ(Eα(xq))
s52: the classification decision of the trained support vector machine on the query sample is as follows:
Figure BDA0003559427200000051
wherein alpha islIs the Lagrangian coefficient, ulFor the first supported sample feature, ylIs it a pairAnd (4) a corresponding label.
The invention has the beneficial effects that:
1. the method utilizes a novel interval supervision contrast loss function to pre-train the model on the base class data set, fixes the parameters in the coder in the pre-trained model, extracts the characteristics of the support image samples in the new class data set and trains an SVM classifier, and finally utilizes the SVM to make classification decision on the query samples.
2. The interval supervision comparison loss function establishes a mathematical model for the contrast relation between the base class samples, instead of only focusing on the class to which the base class samples belong, and the pre-trained backbone network has higher mobility.
3. The supervision contrast loss function in the invention can further reduce the distance of data in the class and increase the distance between classes by increasing the interval parameters, thereby further improving the classification performance.
Drawings
FIG. 1 is a flow chart of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings, so that those skilled in the art can better understand the advantages and features of the present invention, and thus the scope of the present invention is more clearly defined. The embodiments described herein are only a few embodiments of the present invention, rather than all embodiments, and all other embodiments that can be derived by one of ordinary skill in the art without inventive faculty based on the embodiments described herein are intended to fall within the scope of the present invention.
Referring to fig. 1, a small sample image classification method based on interval supervision contrast loss includes the following steps:
Step 1: for a given dataset D, determining a base class dataset D to be processedbaseAnd a new class data set Dnovel(ii) a Wherein
Figure BDA0003559427200000052
xiRepresents the ith base class image sample, yiIndicates its corresponding tag; at DnovelThe N-way-K-shot classification tasks are constructed, and each classification task is provided with a support sample set
Figure BDA0003559427200000053
And querying the sample set
Figure BDA0003559427200000054
Composition of (a) wherein xsDenotes the s-th support sample, ysDenotes the corresponding label, xqRepresents the qth query sample;
step 2: from base class dataset DbaseRandomly decimating M image samples, expanding them to 2M samples using image enhancement techniques, and inputting them to an encoder E with a parameter ααProjector P with parameter betaβIn (c), a set of 2M data features is obtained
Figure BDA0003559427200000055
Assuming any data in the set as an anchor point and taking the anchor point as a positive example sample, taking all samples in the set A (i) which are labeled by the anchor point as positive example samples to construct a positive example sample set P (i), and taking all the rest samples in the set A (i) as negative example samples to construct a negative example sample set N (i);
and step 3: calculating the interval supervision contrast loss between the positive sample set P (i) and the negative sample set N (i), calculating the gradient of the loss function relative to the parameter, and using gradient descent algorithm to the encoder E αParameter α in (v) and projector PβOptimizing the parameter beta in the (-) step;
step 4, pre-training the coder E in the step 3α(. the) is used as a feature extractor to extract features of the support image samples in the new class data set and train the SVM;
and 5: pre-trained encoder E in step 3αAnd the (-) is used as a feature extractor for extracting features of the query image samples in the new class data set, and the trained SVM is used for testing.
Specifically, the specific steps of step 2 are as follows:
s21: from DtrainExtracting M image samples as a training batch image sample set DMFor the ith base class image sample x thereiniEnhanced by enhancement techniques
Figure BDA0003559427200000061
And adding it to the batch image sample set;
s22: data set DMIs input to encoder Eα(. and projector P)β(. o) obtain the corresponding d-dimensional feature vector, where encoder Eα(. to) ResNet12, projector Pβ() is a multi-layer sensor with only one hidden layer;
then the ith base class image sample xiIs characterized by:
ui＝Pβ(Eα(xi))；
s23: through the feature extraction link, a set consisting of 2M data features is obtained at the moment
Figure BDA0003559427200000062
Will uiAs a positive example, the labels and y of all samples in set A (i) iAll samples are regarded as positive samples to form a positive sample set P (i), and the remaining samples are regarded as negative samples to form a negative sample set N (i).
Specifically, the specific steps of step 3 are as follows:
s31: the calculation formula of the interval supervision contrast loss function constructed based on the positive example sample set P (i) and the negative example sample set N (i) is as follows:
Figure BDA0003559427200000063
wherein the content of the first and second substances,
Figure BDA0003559427200000064
in order to be a temperature coefficient of the temperature,
Figure BDA0003559427200000065
is a spacing coefficient, upIs the p-th sample in the positive sample set P (i), uaFor negative sample set N (i), s (-) represents the cosine similarity measure of any two samples, and the expression is:
Figure BDA0003559427200000066
s32: calculating the gradient of the interval supervision contrast loss relative to the parameter, wherein the calculation formula is as follows:
Figure BDA0003559427200000071
wherein
Figure BDA0003559427200000072
Figure BDA0003559427200000073
Figure BDA0003559427200000074
S33: and updating the parameters by using a gradient descent algorithm until the algorithm converges, and finishing the pre-training of the backbone network at the moment.
Specifically, the specific steps of step 4 are as follows:
s41: fixed encoder EαThe parameter alpha in the (-) extracts the feature of the support sample in the new class data set, and the d-dimensional feature of the s-th support sample is expressed as:
us＝Pβ(Eα(xs))
s42: selecting a linear support vector machine to classify and judge the new data samples, wherein an optimized objective function is as follows:
Figure BDA0003559427200000075
s.t.ys(w·us+b)≥1-ξs
ξs≥0,s＝1,2,…,2K
wherein, ysIs a label corresponding to the s-th support sample, C is a penalty coefficient, b is a classification threshold value, and xi sAnd (4) optimizing and solving the above formula to obtain L support vectors, wherein the w is a parameter of the classifier and is a relaxation variable.
Specifically, the specific steps of step 5 are as follows:
s51: fixed encoder EαThe parameter alpha in the (-) extracts the feature of the query sample in the new class data set, and the d-dimensional feature of the qth support sample is expressed as:
uq＝Pβ(Eα(xq))
s52: the classification decision of the trained support vector machine on the query sample is as follows:
Figure BDA0003559427200000076
wherein alpha islIs the Lagrangian coefficient, ulFor the first supported sample feature, ylIs its corresponding tag.
In summary, the invention utilizes a novel interval supervision contrast loss function to pre-train the model on the base class data set, fixes the parameters in the coder in the pre-training model, extracts the features of the support image samples in the new class data set and trains the SVM classifier, and finally utilizes the SVM to make classification decision on the query samples.
The embodiments of the present invention have been described in detail, but the description is only for the preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. All equivalent changes and modifications made within the scope of the present invention shall fall within the scope of the present invention.

Claims (6)

1. A small sample image classification method based on interval supervision contrast loss is characterized by comprising the following steps:
step 1: for a given dataset D, determining a base class dataset D to be processedbaseAnd a new class data set Dnovel(ii) a Wherein
Figure FDA0003559427190000011
xiRepresents the ith base class image sample, yiIndicates its corresponding tag; at DnovelThe N-way-K-shot classification tasks are constructed, and each classification task is provided with a support sample set
Figure FDA0003559427190000012
And querying the sample set
Figure FDA0003559427190000013
Composition of (a) wherein xsDenotes the s-th support sample, ysDenotes the corresponding label, xqRepresents the qth query sample;
step 2: from base class dataset DbaseRandomly decimating M image samples, expanding them to 2M samples using image enhancement techniques, and inputting them to an encoder E with a parameter ααProjector P with parameter betaβIn (c), a set of 2M data features is obtained
Figure FDA0003559427190000014
Assuming any data in the set as an anchor point and taking the anchor point as a positive sample, taking all samples in the set A (i) with the anchor point label as positive samples to construct a positive sample set P (i), and taking all the rest samples in the set A (i) as negative samples to construct a negative sample set N (i);
and step 3: calculating the interval supervision contrast loss between the positive sample set P (i) and the negative sample set N (i), calculating the gradient of the loss function relative to the parameter, and using the gradient Encoder E of reducing algorithm pairαMiddle parameter alpha and projector PβOptimizing the parameter beta in the (-) step;
step 4, pre-training the coder E in the step 3α(. the) is used as a feature extractor to extract features of the support image samples in the new class data set and train the SVM;
and 5: pre-trained encoder E in step 3αAnd the (-) is used as a feature extractor for extracting features of the query image samples in the new class data set, and the trained SVM is used for testing.
2. The method for classifying small sample images based on interval supervision contrast loss according to claim 1 is characterized in that the specific steps of step 1 are as follows:
s11: for a given data set D, it is partitioned into three subsets, the training set DtrainVerification set DvalAnd test set DtestThe classification categories in each subset are not the same, namely:
Figure FDA0003559427190000015
Dtrain∪Dval∪Dtest＝D；
s12: will train set DtrainAs base class dataset DbaseTest set DtestAs a new class data set Dnovel(ii) a At DnovelRandomly extracting N classes in the data set, and randomly extracting K samples in each class to obtain a support sample set
Figure FDA0003559427190000016
Extracting a batch of samples from the residual data in the N categories to obtain a query sample set
Figure FDA0003559427190000017
3. The method for classifying small sample images based on interval supervision contrast loss according to claim 1 is characterized in that the specific steps of the step 2 are as follows:
S21: from DtrainExtracting M image samples as a training batch image sample set DMFor the ith base class image sample x thereiniEnhanced by enhancement techniques
Figure FDA0003559427190000018
And adding it to the batch image sample set;
s22: data set DMIs input to encoder Eα(. and projector P)β(. o) obtain the corresponding d-dimensional feature vector, where encoder Eα(. to) ResNet12, projector Pβ() is a multi-layer sensor with only one hidden layer;
then the ith base class image sample xiIs characterized by:
ui＝Pβ(Eα(xi))；
s23: through the feature extraction link, a set consisting of 2M data features is obtained at the moment
Figure FDA0003559427190000021
Will uiAs a positive example, the labels and y of all samples in set A (i)iAll samples are regarded as positive samples to form a positive sample set P (i), and the remaining samples are regarded as negative samples to form a negative sample set N (i).
4. The method for classifying small sample images based on interval supervision contrast loss according to claim 1 is characterized in that the specific steps of the step 3 are as follows:
s31: the calculation formula of the interval supervision contrast loss function constructed based on the positive example sample set P (i) and the negative example sample set N (i) is as follows:
Figure FDA0003559427190000022
wherein the content of the first and second substances,
Figure FDA0003559427190000023
In order to be a temperature coefficient of the temperature,
Figure FDA0003559427190000024
is a spacing coefficient, upIs the p sample in the positive example sample set P (i), uaFor negative sample set N (i), s (-) represents the cosine similarity measure of any two samples, and the expression is:
Figure FDA0003559427190000025
s32: calculating the gradient of the interval supervision contrast loss relative to the parameter, wherein the calculation formula is as follows:
Figure FDA0003559427190000026
wherein
Figure FDA0003559427190000027
Figure FDA0003559427190000028
Figure FDA0003559427190000029
S33: and updating the parameters by using a gradient descent algorithm until the algorithm converges, and finishing the pre-training of the backbone network at the moment.
5. The method for classifying small sample images based on interval supervision contrast loss according to claim 1 is characterized in that the specific steps of the step 4 are as follows:
s41: fixed encoder EαThe parameter alpha in the (-) extracts the feature of the support sample in the new class data set, and the d-dimensional feature of the s-th support sample is expressed as:
us＝Pβ(Eα(xs))
s42: selecting a linear support vector machine to classify and judge the new data samples, wherein an optimized objective function is as follows:
Figure FDA0003559427190000031
s.t.ys(w·us+b)≥1-ξs
ξs≥0,s＝1,2,…,2K
wherein, ysIs a label corresponding to the s-th support sample, C is a penalty coefficient, b is a classification threshold value, and xisAnd (4) optimizing and solving the above formula to obtain L support vectors, wherein the w is a parameter of the classifier and is a relaxation variable.
6. The method for classifying small sample images based on interval supervision contrast loss according to claim 1 is characterized in that the specific steps of the step 5 are as follows:
S51: fixed encoder EαThe parameter alpha in the (-) extracts the feature of the query sample in the new class data set, and the d-dimensional feature of the qth support sample is expressed as:
uq＝Pβ(Eα(xq))
s52: the classification decision of the trained support vector machine on the query sample is as follows:
Figure FDA0003559427190000032
wherein alpha islIs the Lagrangian coefficient, ulFor the first supported sample feature, ylIs its corresponding tag.
CN202210289086.9A 2022-03-22 2022-03-22 Small sample image classification method based on interval supervision contrast loss Withdrawn CN114580566A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210289086.9A CN114580566A (en) 2022-03-22 2022-03-22 Small sample image classification method based on interval supervision contrast loss

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210289086.9A CN114580566A (en) 2022-03-22 2022-03-22 Small sample image classification method based on interval supervision contrast loss

Publications (1)

Publication Number Publication Date
CN114580566A true CN114580566A (en) 2022-06-03

Family

ID=81783167

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210289086.9A Withdrawn CN114580566A (en) 2022-03-22 2022-03-22 Small sample image classification method based on interval supervision contrast loss

Country Status (1)

Country Link
CN (1) CN114580566A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115203420A (en) * 2022-07-25 2022-10-18 腾讯科技（深圳）有限公司 Entity relationship classification model training method, entity relationship classification method and device
CN115294386A (en) * 2022-07-06 2022-11-04 南通大学 Image classification method based on regularization supervision loss function
CN115984621A (en) * 2023-01-09 2023-04-18 宁波拾烨智能科技有限公司 Small sample remote sensing image classification method based on restrictive prototype comparison network
CN116168257A (en) * 2023-04-23 2023-05-26 安徽大学 Small sample image classification method, device and storage medium based on sample generation

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115294386A (en) * 2022-07-06 2022-11-04 南通大学 Image classification method based on regularization supervision loss function
CN115294386B (en) * 2022-07-06 2023-11-24 南通大学 Image classification method based on regularization supervision loss function
CN115203420A (en) * 2022-07-25 2022-10-18 腾讯科技（深圳）有限公司 Entity relationship classification model training method, entity relationship classification method and device
CN115203420B (en) * 2022-07-25 2024-04-26 腾讯科技（深圳）有限公司 Entity relationship classification model training method, entity relationship classification method and device
CN115984621A (en) * 2023-01-09 2023-04-18 宁波拾烨智能科技有限公司 Small sample remote sensing image classification method based on restrictive prototype comparison network
CN115984621B (en) * 2023-01-09 2023-07-11 宁波拾烨智能科技有限公司 Small sample remote sensing image classification method based on restrictive prototype comparison network
CN116168257A (en) * 2023-04-23 2023-05-26 安徽大学 Small sample image classification method, device and storage medium based on sample generation

Similar Documents

Publication Publication Date Title
CN110298037B (en) Convolutional neural network matching text recognition method based on enhanced attention mechanism
Ruby et al. Binary cross entropy with deep learning technique for image classification
Wu et al. Unsupervised feature learning via non-parametric instance discrimination
Deng et al. Multitask emotion recognition with incomplete labels
Savchenko Efficient facial representations for age, gender and identity recognition in organizing photo albums using multi-output ConvNet
CN114580566A (en) Small sample image classification method based on interval supervision contrast loss
WO2022037233A1 (en) Small sample visual target identification method based on self-supervised knowledge transfer
Biswas et al. Simultaneous active learning of classifiers & attributes via relative feedback
CN112614538A (en) Antibacterial peptide prediction method and device based on protein pre-training characterization learning
Duong et al. Shrinkteanet: Million-scale lightweight face recognition via shrinking teacher-student networks
Ju et al. Fish species recognition using an improved AlexNet model
US20210319215A1 (en) Method and system for person re-identification
Ostyakov et al. Label denoising with large ensembles of heterogeneous neural networks
CN113554110B (en) Brain electricity emotion recognition method based on binary capsule network
CN113361278B (en) Small sample named entity identification method based on data enhancement and active learning
CN113672718B (en) Dialogue intention recognition method and system based on feature matching and field self-adaption
WO2021128704A1 (en) Open set classification method based on classification utility
Xu et al. Weakly supervised facial expression recognition via transferred DAL-CNN and active incremental learning
Okokpujie et al. Predictive modeling of trait-aging invariant face recognition system using machine learning
CN112434686B (en) End-to-end misplaced text classification identifier for OCR (optical character) pictures
CN116935411A (en) Radical-level ancient character recognition method based on character decomposition and reconstruction
CN116994092A (en) K-based mutual neighbor pseudo tag screening method
He et al. Addressing the Overfitting in Partial Domain Adaptation with Self-Training and Contrastive Learning
CN112200260B (en) Figure attribute identification method based on discarding loss function
CN113297376A (en) Legal case risk point identification method and system based on meta-learning

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
WW01 Invention patent application withdrawn after publication

Application publication date: 20220603

WW01 Invention patent application withdrawn after publication