Active domain adaptive image classification method and system based on clustering environment perception

Abstract

The application discloses a kind of active field adaptive image classification method and system based on clustering environment perception, simultaneously consider individual and the information value of cluster environment where it is, to select several images to be labeled and join training sample;When selecting unlabeled image, the information of image dataset is calculated by single point and its cluster environment, to measure whether the image has higher labeling value;Meanwhile, in the image sample selection process, consider the representative strategy of image dataset to select representative and different labeled distribution image dataset with labeled image dataset, ensure that image dataset with different labeled distribution is selected, to provide more novel information to train model, and explore the image information most valuable to the current labeled distribution model improvement, to improve the generalization ability of model. Through learning the distribution information of unlabeled image data, model can more accurately carry out classification prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, more particularly, to an active domain adaptive image classification method and system based on cluster environment perception. BACKGROUND

[0002] In actual scenarios, it is usually feasible to obtain labels for a small part of the target data, but in order to save costs, it is better to select a subset of the most valuable data through active learning. Active learning is an important learning paradigm in machine learning, and its purpose is to train an effective model with as few query samples as possible.

[0003] A key problem of existing active domain adaptive algorithms is that they only consider the information value of individual points or the information value of the cluster environment in which the individual points are located, which may lead to the selection of low-value samples. When only considering individual points, due to the lack of understanding of the cluster environment in which they are located, outliers may be selected because outliers also contain high-value individual information for the model. In addition, this also leads to a lack of understanding of whether the samples are in a tight cluster environment or a more dispersed environment. The cluster environment to which the samples belong should be considered when selecting samples to better judge the information value of the data. In order to solve this problem, we propose to integrate the information value of individual points and the information value of the cluster environment into the sample selection process; by considering both aspects, more accurate classification prediction can be made.

[0004] The prior art provides an image classification method, system, device and medium based on active domain adaptation, the method comprising: initializing an annotated data set and an unannotated data set according to the obtained source domain image data set and target domain image data set, and initializing and training a preset deep neural network model according to the annotated data set to obtain a target domain image classification model, then establishing an information scoring model and an information sampling model according to the annotated data set, the unannotated data set and the target domain image classification model, and performing active domain adaptive learning iteration training on the target domain image classification model, updating the annotated data set and the unannotated data set until a preset iteration stopping condition is reached, and inputting a to-be-classified image into the target domain image classification model for classification prediction to obtain an image classification result. The present application designs a sampling and training strategy based on the heterogeneity of sample differences, effectively improves the classification accuracy and generalization ability of the model, and has strong universality. SUMMARY

[0005] The application provides a cluster environment perception-based active field adaptive image classification method and system, integrates information values of individual points and information values of a cluster environment into a sample selection process, thereby improving sample selection quality, improving model performance and generalization ability, effectively dealing with data scarcity and high annotation cost, and more accurately performing classification prediction.

[0006] To solve the above technical problems, the technical scheme of the application is as follows:

[0007] The application provides a cluster environment perception-based active field adaptive image classification method, which comprises the following steps:

[0008] S1: acquiring a labeled source domain image dataset and an unlabeled target domain image dataset;

[0009] S2: pre-processing the labeled source domain image dataset and the unlabeled target domain image dataset to obtain a pre-processed source domain labeled image dataset and an unlabeled target domain image dataset, and combining the pre-processed source domain labeled image dataset and the unlabeled target domain image dataset to form a training image dataset;

[0010] S3: performing adversarial training on a constructed classification model by using the training image dataset to obtain a pre-trained classification model;

[0011] S4: inputting the pre-processed unlabeled target domain image dataset into the pre-trained classification model to calculate a single-point information amount score and a cluster information amount score of each pre-processed unlabeled target domain image;

[0012] S5: calculating a sample total information amount score of each pre-processed unlabeled target domain image according to the single-point information amount score and the cluster information amount score of the pre-processed unlabeled target domain image;

[0013] S6: based on the sample information amount score, screening the pre-processed unlabeled target domain image, labeling the screened unlabeled target domain image, and obtaining a labeled target domain image dataset; and adding the labeled target domain image dataset to the training image dataset;

[0014] S7: repeating steps S3-S6, setting a total loss function, and repeating until a preset adversarial training number is reached to obtain a trained classification model;

[0015] S8: acquiring an image to be classified, inputting the image to be classified into the trained classification model to perform classification prediction, and obtaining an image classification result.

[0016] Preferably, in S2, the specific method for pre-processing is as follows:

[0017] The source domain labeled image dataset and the target domain unlabeled image dataset are normalized, and random data augmentation is used to obtain the preprocessed source domain labeled image dataset and the target domain unlabeled image dataset.

[0018] Preferably, in S4, the preprocessed target domain unlabeled image dataset is input into the pre-trained classification model to calculate the single-point information quantity score and the cluster information quantity score of each preprocessed unlabeled target domain image.

[0019] S41: the pre-trained classification model, wherein the classification model comprises a feature extractor and a classifier connected in sequence; the classifier comprises a domain classifier and a category classifier arranged side by side;

[0020] S42: performing feature extraction on the preprocessed target domain unlabeled image data by using the feature extractor to obtain an information feature vector;

[0021] S43: the feature vector is calculated by the domain classifier and the category classifier to obtain the single-point information quantity score and the cluster information quantity score.

[0022] Preferably, in S43, the single-point information quantity score is:

[0023] ISP(x i )=(1-U(G f (x i )))+αV(G f (x i ))

[0024] Wherein, U is the uncertainty score of the unlabeled target domain image data; V is the target domain proximity score of the unlabeled target domain image data; and a is the balance relationship between the uncertainty and the target domain proximity score.

[0025] Preferably, in S43, the cluster information quantity score is:

[0026] ISC(x i )=CIC(x i )*CIK(x i )

[0027] Wherein, ISC(x i ) is the cluster information quantity score; CIC(x i ) is the cluster cohesion; and CIK(x i ) is the cluster internal information quantity.

[0028] Preferably, the cluster cohesion is:

[0029]

[0030] The cluster internal information quantity is:

[0031]

[0032] Q ij represents the jth unlabeled target domain image data of x i ; Dis(·|·) represents the measured distance between two unlabeled target domain data; G f is a feature extractor; K is the amount of unlabeled target domain image data.

[0033] Preferably, in S5, the sample total information quantity score is:

[0034] ISPC(x i )=ISP(x i )+ISC(x i )

[0035] Where ISPC(x i ) is the total information score; ISP(x i ) is the single-point information quantity score; ISC(x i ) is the cluster information quantity score; x i is the information feature vector.

[0036] Preferably, in S6, based on the sample information quantity score, the preprocessed target domain unlabeled image is screened, the screened target domain unlabeled image is labeled, and a labeled target domain image dataset is obtained. Specifically:

[0037] According to the sample total information quantity score, the target domain unlabeled image data is sorted from large to small, redundant target domain unlabeled image data is removed according to a preset rule, a certain number of target domain unlabeled image data sorted at the top is selected based on the remaining target domain unlabeled image dataset for manual labeling, it is ensured that the selected certain number of target domain unlabeled image data is manually labeled, and a labeled target domain image dataset is obtained.

[0038] Preferably, in S7, the total loss function is:

[0039]

[0040] Where L adv is an alignment loss function; L im is an information maximization strategy loss function; L c is a cross-entropy loss function in the labeled source domain and target domain dataset.

[0041] The cross-entropy loss function is:

[0042]

[0043] wherein G c is a class classifier; G f is a feature extractor; D S represents a source domain dataset; D LT is a labeled target domain dataset; x s represents source domain data; y s represents a label in source domain labeled data; x t represents target domain data; y t represents a label in target domain labeled data;

[0044] The alignment loss function is:

[0045]

[0046] wherein G d is a domain classifier, D UT is an unlabeled target domain dataset;

[0047] The information maximization strategy loss function is:

[0048]

[0049] wherein C represents the number of sample categories; δ c represents the cth element in the C-dimensional vector output; D KL represents the KL divergence; represents the average output of G c over the entire unlabeled target domain data; l C is a first vector.

[0050] The application also provides an active domain adaptive image classification system based on clustering environment perception, which is used to implement the above method, and the system comprises:

[0051] a data acquisition module, which acquires a labeled source domain image dataset and an unlabeled target domain image dataset;

[0052] a preprocessing module, which pre-processes the labeled source domain image dataset and the unlabeled target domain image dataset to obtain a pre-processed source domain labeled image dataset and a target domain unlabeled image dataset, and combines the pre-processed source domain labeled image dataset and the target domain unlabeled image dataset to form a training image dataset;

[0053] a pre-training module, which performs adversarial training on a constructed classification model by using the training image dataset to obtain a pre-trained classification model;

[0054] The single-point and cluster information quantity score module inputs the preprocessed target domain unlabeled image dataset into the pre-trained classification model, and calculates the single-point information quantity score and the cluster information quantity score of each preprocessed unlabeled target domain image.

[0055] The total information quantity score module calculates the sample total information quantity score of the preprocessed unlabeled target domain image according to the single-point information quantity score and the cluster information quantity score of each preprocessed unlabeled target domain image.

[0056] The screening and labeling module screens the preprocessed unlabeled target domain image based on the sample information quantity score, labels the screened unlabeled target domain image, and obtains a labeled target domain image dataset; and adds the labeled target domain image dataset into the training image dataset.

[0057] The updating module is configured to obtain the training image dataset, return the single-point and cluster information quantity score module, set a total loss function, and obtain the trained classification model until a preset number of adversarial training is reached.

[0058] The classification module obtains a to-be-classified image, inputs the to-be-classified image into the trained classification model for classification prediction, and obtains an image classification result.

[0059] Compared with the prior art, the technical scheme of the present application has the following beneficial effects:

[0060] The present application simultaneously considers the information value of an individual and its clustering environment to select a plurality of images for labeling and adding to the training sample. In the selection of unlabeled images, the present application calculates the information of the image dataset through the single point and its cluster environment to measure whether the image has high labeling value. At the same time, in the image sample selection process, the representative strategy of the image dataset is considered to select the image dataset with representative and different labeling distribution from the labeled image dataset, to ensure the selection of the image dataset with different labeling distribution, to provide more novel information to train the model, and to explore the image information most valuable to the improvement of the current labeling distribution model, thereby improving the generalization ability of the model. Through learning the distribution information of the unlabeled data, the model can better understand the entire data space and more accurately perform classification prediction. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 The flowchart of the active domain self-adaptive image classification method based on cluster environment perception described in embodiment 1;

[0062] Figure 2 The structural diagram of the screened target domain unlabeled image described in embodiment 2;

[0063] Figure 3A structural schematic diagram of a clustering environment perception based active domain adaptive image classification system described in embodiment 3. DETAILED DESCRIPTION

[0064] The drawings are only used for illustrative description and cannot be understood as limiting the patent;

[0065] In order to better illustrate the embodiments, some components in the drawings may be omitted, enlarged or reduced, and do not represent the actual size of the product;

[0066] For those skilled in the art, it is understandable that some well-known structures in the drawings and their descriptions may be omitted.

[0067] The technical solutions of the present application will be further described below in combination with the drawings and embodiments.

[0068] Embodiment 1

[0069] The present embodiment provides a clustering environment perception based active domain adaptive image classification method, as shown in Figure 1 The method comprises the following steps:

[0070] S1: obtaining a labeled source domain image dataset and an unlabeled target domain image dataset;

[0071] S2: preprocessing the labeled source domain image dataset and the unlabeled target domain image dataset to obtain a preprocessed source domain labeled image dataset and an unlabeled target domain image dataset, and forming a training image dataset;

[0072] S3: using the training image dataset to perform adversarial training on a constructed classification model to obtain a pre-trained classification model;

[0073] S4: inputting the preprocessed unlabeled target domain image dataset into the pre-trained classification model to calculate the single-point information quantity score and the clustering information quantity score of each preprocessed unlabeled target domain image;

[0074] S5: calculating the total sample information quantity score of the preprocessed unlabeled target domain image according to the single-point information quantity score and the clustering information quantity score of each preprocessed unlabeled target domain image;

[0075] S6: based on the sample information quantity score, screening the preprocessed unlabeled target domain image, labeling the screened unlabeled target domain image to obtain a labeled target domain image dataset; adding the labeled target domain image dataset to the training image dataset;

[0076] S7: repeating steps S3-S6, setting a total loss function, until a preset adversarial training number is reached, to obtain a trained classification model;

[0077] S8: Obtain an image to be classified, input the image to be classified into the trained classification model for classification prediction, and obtain an image classification result.

[0078] Embodiment 2

[0079] The application provides an active domain adaptive image classification method based on cluster environment perception, and the method comprises the following steps:

[0080] S1: Obtain a labeled source domain image dataset and an unlabeled target domain image dataset;

[0081] S2: Preprocess the labeled source domain image dataset and the unlabeled target domain image dataset to obtain a preprocessed source domain labeled image dataset and an unlabeled target domain image dataset, and combine the two datasets to form a training image dataset;

[0082] The specific method for preprocessing is as follows:

[0083] The source domain labeled image dataset and the target domain unlabeled image dataset are subjected to normalization processing and random data enhancement to obtain the preprocessed source domain labeled image dataset and the target domain unlabeled image dataset.

[0084] S3: Perform adversarial training on a constructed classification model by using the training image dataset to obtain a pre-trained classification model;

[0085] S4: Input the preprocessed target domain unlabeled image dataset into the pre-trained classification model to calculate a single-point information quantity score and a cluster information quantity score of each preprocessed unlabeled target domain image;

[0086] As shown in the formula, the preprocessed target domain unlabeled image dataset is input into the pre-trained classification model to calculate a single-point information quantity score and a cluster information quantity score of each preprocessed unlabeled target domain image, and the calculation is specifically as follows: Figure 2

[0087] S41: The pre-trained classification model comprises a feature extractor and a classifier connected in sequence; the classifier comprises a domain classifier and a category classifier arranged side by side;

[0088] S42: The feature extractor is used to extract features of the preprocessed target domain unlabeled image dataset to obtain an information feature vector;

[0089] S43: The feature vector is calculated by the domain classifier and the category classifier to obtain a single-point information quantity score and a cluster information quantity score;

[0090] In S43, the single-point information quantity score is as follows:

[0091] ​ISP(x i )=(1-U(G f (x i )))+αV(G f (x i ))

[0092] wherein U is an uncertainty score of the unlabeled target domain image data; V is a target domain proximity score of the unlabeled target domain image data; and a is a balance between the uncertainty and the target domain proximity score;

[0093] The cluster information score is:

[0094] ISC(x i )=CIC(x i )*CIK(x i )

[0095] wherein ISC(x i ) is a cluster information score; CIC(x i ) is a cluster cohesion; and CIK(x i ) is a cluster internal information;

[0096] The cluster cohesion is:

[0097]

[0098] The cluster internal information is:

[0099]

[0100] wherein Q ij represents the jth unlabeled target domain image data of x i ; Dis(·|·) represents a measured distance between two unlabeled target domain data; G f is a feature extractor; and K is an amount of unlabeled target domain image data.

[0101] S5: calculating a sample total information score of each pre-processed unlabeled target domain image according to the single-point information score and the cluster information score of the pre-processed unlabeled target domain image;

[0102] The sample total information score is:

[0103] ISPC(x i )=ISP(x i )+ISC(x i )

[0104] wherein ISPC(x i ) is a total information score; ISP(x i) is a single point information quantity score; ISC(x i ) is a cluster information quantity score; x i is an information feature vector.

[0105] S6: Based on the sample information quantity score, the pre-processed unlabeled target domain image is screened, and the screened unlabeled target domain image is labeled to obtain a labeled target domain image dataset; the labeled target domain image dataset is added to the training image dataset;

[0106] Based on the sample information quantity score, the pre-processed target domain unlabeled image is screened, and the screened target domain unlabeled image is labeled to obtain a labeled target domain image dataset, which is specifically:

[0107] According to the total sample information quantity score, the target domain unlabeled image data is sorted from large to small, the redundant target domain unlabeled image data is removed according to the representative strategy, and the remaining target domain unlabeled image dataset is selected. A number of target domain unlabeled image data with high ranking are manually labeled to ensure that the selected target domain unlabeled image data are manually labeled to obtain a labeled target domain image dataset;

[0108] The redundant target domain unlabeled image data is removed according to the following method:

[0109] Point distance release: In order to keep the sampling points at a certain distance to ensure their diversity, for a candidate instance x k , we check whether its nearest neighbor has been selected. If so, it means that a sample with similar features to the candidate sample has been labeled, so we skip this sample to avoid redundancy.

[0110] Information score difference exclusion: In order to ensure the effectiveness of the selected samples, for each candidate instance x k , we calculate the difference between its ISC and ISP values. When the difference is large, it means that there is a large gap between the score of the point and its clustering environment, which means that the point cannot well represent the nearby points, i.e. lack of representation, so this point is skipped.

[0111] S7: Repeat steps S3-S6, set the total loss function, until the preset number of adversarial training is reached, and obtain a trained classification model;

[0112] The total loss function is:

[0113]

[0114] Where, L adv is an alignment loss function; L im is an information maximization strategy loss function; Lc is a cross-entropy loss function in the labeled source domain and target domain data sets;

[0115] The cross-entropy loss function is:

[0116]

[0117] where G c is a class classifier; G f is a feature extractor; D S represents a source domain data set; D LT is a labeled target domain data set; x s represents source domain data; y s represents a label in the source domain labeled data; x t represents target domain data; y t represents a label in the target domain labeled data;

[0118] The alignment loss function is:

[0119]

[0120] where G d is a domain classifier, D UT is an unlabeled target domain data set;

[0121] The information maximization strategy loss function is:

[0122]

[0123] where C represents the number of sample categories; δ c represents the cth element in the C-dimensional vector output; D KL represents the KL divergence; represents the average output of G c over the entire unlabeled target domain data; l C is a first vector.

[0124] S8: Obtain an image to be classified, input the image to be classified into the trained classification model for classification prediction, and obtain an image classification result.

[0125] Embodiment 3

[0126] This embodiment also provides an active domain adaptive image classification system based on clustering environment perception, which is used to implement the method of embodiments 1 or 2, as shown in Figure 3 The system comprises:

[0127] A data acquisition module acquires a labeled source domain image data set and an unlabeled target domain image data set.

[0128] The pre-processing module pre-processes the labeled source domain image dataset and the unlabeled target domain image dataset to obtain a pre-processed source domain labeled image dataset and a target domain unlabeled image dataset, and combines the pre-processed source domain labeled image dataset and the target domain unlabeled image dataset to form a training image dataset;

[0129] The pre-training module performs adversarial training on the constructed classification model by using the training image dataset to obtain a pre-trained classification model.

[0130] The single point and cluster information quantity score module inputs the pre-processed target domain unlabeled image dataset into the pre-trained classification model, and calculates a single point information quantity score and a cluster information quantity score of each pre-processed target domain unlabeled image.

[0131] The total information quantity score module calculates a total information quantity score of each pre-processed target domain unlabeled image according to the single point information quantity score and the cluster information quantity score of the pre-processed target domain unlabeled image.

[0132] The screening and labeling module screens the pre-processed target domain unlabeled image based on the sample information quantity score, labels the screened target domain unlabeled image, and obtains a labeled target domain image dataset; and adds the labeled target domain image dataset into the training image dataset.

[0133] The updating module is configured to obtain the training image dataset, return to the single point and cluster information quantity score module, set a total loss function, and obtain a trained classification model until a preset adversarial training number is reached.

[0134] The classification module obtains an image to be classified, inputs the image to be classified into the trained classification model for classification prediction, and obtains an image classification result.

[0135] The same or similar reference signs correspond to the same or similar components;

[0136] The terms used to describe the positional relationship in the drawings are only used for illustrative description, and should not be understood as a limitation on the patent;

[0137] Obviously, the above embodiments of the present application are only examples for clearly illustrating the present application, and are not intended to limit the embodiments of the present application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, it is not necessary or possible to exhaust all the embodiments. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the claims of the present application.

Claims

1. A method for active domain adaptive image classification based on clustering environment perception, characterized in that, The method comprises: S1: obtaining a labeled source domain image dataset and an unlabeled target domain image dataset; S2: preprocessing the labeled source domain image dataset and the unlabeled target domain image dataset to obtain a preprocessed source domain labeled image dataset and a target domain unlabeled image dataset, and combining the preprocessed source domain labeled image dataset and the target domain unlabeled image dataset to form a training image dataset; S3: performing adversarial training on a constructed classification model by using the training image dataset to obtain a pre-trained classification model; S4: inputting the preprocessed target domain unlabeled image dataset into the pre-trained classification model to calculate a single-point information quantity score and a cluster information quantity score of each preprocessed unlabeled target domain image, wherein an expression of the single-point information quantity score is: ISP(x i ) = (1 - U(G f (x i )) ) + aV(G f (x i )) wherein U is an uncertainty score for the unlabeled target domain image data; V is a target domain closeness score for the unlabeled target domain image data; G f is a feature extractor; a is a balancing relationship between the uncertainty and target domain closeness scores; an expression of the cluster information quantity score is: ISC(x i ) = CIC(x i ) * CIK(x i ) where ISC(x i ) is the cluster information score; CIC(x i ) is the cluster cohesion; and CIK(x i ) is the cluster internal information. S5: calculating a sample total information quantity score of the preprocessed unlabeled target domain image according to the single-point information quantity score and the cluster information quantity score of the preprocessed unlabeled target domain image; S6: based on the sample total information quantity score, screening the preprocessed target domain unlabeled image, labeling the screened target domain unlabeled image, and obtaining a labeled target domain image dataset; and adding the labeled target domain image dataset to the training image dataset; S7: repeating steps S3-S6, setting a total loss function, and repeating until a preset adversarial training number is reached to obtain a trained classification model; S8: obtaining an image to be classified, inputting the image to be classified into the trained classification model for classification prediction, and obtaining an image classification result.

2. The method of claim 1, wherein, In S2, the specific method of preprocessing is: The source domain labeled image dataset and the target domain unlabeled image dataset are normalized and random data augmentation is used to obtain the preprocessed source domain labeled image dataset and the target domain unlabeled image dataset.

3. The method of claim 1, wherein the method further comprises: In S4, inputting the preprocessed target domain unlabeled image dataset into the pre-trained classification model to calculate the single-point information quantity score and the cluster information quantity score of each preprocessed unlabeled target domain image is specifically: S41: the pre-trained classification model, wherein the classification model comprises a feature extractor and a classifier connected in sequence; the classifier comprises a domain classifier and a category classifier arranged side by side; S42: using the feature extractor to extract features of the preprocessed target domain unlabeled image data to obtain an information feature vector; S43: the feature vector is calculated by the domain classifier and the category classifier to obtain the single-point information quantity score and the cluster information quantity score.

4. The method of claim 1, wherein the method further comprises: The cluster cohesion is: The cluster internal information quantity is: wherein Q ij represents x i the jthunlabeled target domain image data; Dis(·|·) represents the measured distance between two unlabeled target domain data; G f is a feature extractor; K is the amount of unlabeled target domain image data.

5. The method of claim 1, wherein, In S5, the sample total information quantity score is: ISPC(x i ) = ISP(x i ) + ISC(x i ) where ISPC(x i ) is the total information score; ISP(x i ) is the single point information score; ISC(x i ) is the cluster information score; and x i is the information feature vector.

6. The method of claim 1, wherein, In S6, based on the sample information quantity score, the preprocessed target domain unlabeled image is screened, the screened target domain unlabeled image is labeled, and a labeled target domain image dataset is obtained, which is specifically: According to the target domain unlabeled image data corresponding to the sample total information amount score in descending order, the redundant target domain unlabeled image data is removed according to a preset rule, and a plurality of target domain unlabeled image data with high ranking is selected based on the remaining target domain unlabeled image data set to be manually labeled, so as to ensure that the selected target domain unlabeled image data is manually labeled, and a labeled target domain image data set is obtained.

7. The method of claim 1, wherein, In S7, the total loss function is: where L adv is the alignment loss function; L im is the information maximization policy loss function; L c is the cross-entropy loss function in the labeled source and target domain datasets; The cross-entropy loss function is: where G c is a class classifier; G f is a feature extractor; D S represents a source domain dataset; D LT is a labeled target domain dataset; x s represents source domain data; y s represents a label in source domain labeled data; x t represents target domain data; y t represents a label in target domain labeled data; The alignment loss function is: where G d is a domain classifier, D UT is an unlabeled target domain dataset; The information maximization strategy loss function is: where C represents the number of sample categories; δ c represents the cth element in the C-dimensional vector output; D KL represents the KL divergence; represents G c the average output over the entire unlabeled target domain data; l C is the first vector.

8. A cluster-based context-aware active domain adaptive image classification system for implementing the method of any one of claims 1-7, characterized by, The system comprises: A data acquisition module acquires a labeled source domain image data set and an unlabeled target domain image data set. A preprocessing module preprocesses the labeled source domain image data set and the unlabeled target domain image data set to obtain a preprocessed source domain labeled image data set and a target domain unlabeled image data set, and forms a training image data set. A pre-training module performs adversarial training on a constructed classification model using the training image data set to obtain a pre-trained classification model. A single-point and cluster information amount score module inputs the preprocessed target domain unlabeled image data set into the pre-trained classification model, calculates the single-point information amount score and the cluster information amount score of each preprocessed unlabeled target domain image. A total information amount score module calculates the sample total information amount score of each preprocessed unlabeled target domain image according to the single-point information amount score and the cluster information amount score of each preprocessed unlabeled target domain image. A screening and labeling module screens the preprocessed unlabeled target domain image based on the sample information amount score, labels the screened unlabeled target domain image, and obtains a labeled target domain image data set; and adds the labeled target domain image data set to the training image data set. An update module is configured to acquire the training image data set, return to the single-point and cluster information amount score module, set the total loss function, and obtain a trained classification model until a preset adversarial training number is reached. A classification module acquires an image to be classified, inputs the image to be classified into the trained classification model for classification prediction, and obtains an image classification result.

Images