A general domain adaptation image classification model implementation method and system
By training closed-set and open-set classifiers, using the maximum open-set entropy to separate samples and combining a neighbor aggregation strategy, the problems of generalized adaptive mid-domain offset and label difference are solved, thus improving the generalization ability of the image classification model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGZHOU UNIVERSITY
- Filing Date
- 2023-09-19
- Publication Date
- 2026-05-19
AI Technical Summary
Existing general domain adaptation methods struggle to set reasonable thresholds when dealing with domain and label differences, leading to severe domain shift problems. Furthermore, they fail to adequately generalize the ability of closed-set spaces to measure sample uncertainty.
The closed-set and open-set classifiers are trained using the cross-entropy loss function. The confidence level of a sample is determined by the maximum open-set entropy. The most easily confused categories are separated and a transition zone is set. The simple and difficult samples are separated by the neighbor aggregation strategy. Batch training is performed by combining cross-entropy, selection optimization strategy and neighbor aggregation loss function to obtain the final general domain adaptive image classification model.
It effectively measures the uncertainty of target samples, separates shared samples from unknown samples, improves the model's generalization ability in different domains, and solves the problems of domain shift and negative transfer of private samples in the target domain.
Smart Images
Figure CN117253105B_ABST
Abstract
Description
Technical Field
[0001] This document relates to the field of computer technology, and in particular to a method and system for implementing a general domain adaptive image classification model. Background Technology
[0002] In recent years, deep neural networks have made significant progress in image recognition. However, most excellent image classification techniques require large amounts of annotated training data, leading to expensive manual annotation costs. To alleviate this burden, unsupervised domain adaptation (UDA) has emerged. The goal of this technique is to transfer knowledge learned by a neural network from a labeled auxiliary domain (source domain) to a new, unlabeled domain (target domain). Although UDA has achieved good performance, a limitation is its assumption that the source and target domains share the same label set. This is often not true in real-world scenarios; in fact, we usually cannot obtain prior knowledge about the relationship between the label sets of the two domains. To address this issue, subsequent researchers proposed the Universal Domain Adaptation (UniDA) technique to handle general scenarios where both domain and label differences exist. Currently, commonly used UniDA methods use various criteria in the closed set space to measure the uncertainty of the target sample, such as the closed set entropy of the sample, and set a fixed threshold to distinguish "unknown" samples in the target domain. At the same time, they force domain alignment in shared classes with a large number of noisy samples. However, these methods ignore the limitation of the poor generalization ability of the closed set space to measure sample uncertainty, making it difficult to set a reasonable threshold, which further aggravates the domain offset problem. Summary of the Invention
[0003] The purpose of this invention is to provide a method and system for implementing a general domain adaptive image classification model, aiming to solve the above-mentioned problems in the prior art.
[0004] This invention provides a method for implementing a general domain adaptive image classification model, including:
[0005] Set the source domain image dataset with classification labels and the target domain image dataset with a different distribution than the source domain data and no classification labels as the image data and classification labels for training the classification model;
[0006] Based on the source domain image dataset, the closed-set classifier and open-set classifier in the classification model are trained using the cross-entropy loss function. The closed-set classifier outputs the classification probability of a sample belonging to each category in the source domain. The multiple sub-binary classifiers of the open-set classifier output the intra-class and out-class classification probabilities of a sample belonging to the corresponding category in the source domain, as well as the maximum open-set entropy of the sample. The maximum open-set entropy is used to determine the confidence level of the sample.
[0007] The most easily confused category of the target domain image dataset samples is determined based on the maximum open set entropy, and a transition zone is set in the most easily confused category. An optimization strategy is selected based on the maximum open set entropy to separate simple samples outside the transition zone with confidence levels higher than a preset first threshold, and a neighbor aggregation strategy is combined to separate difficult samples within the transition zone with confidence levels lower than a preset second threshold.
[0008] Based on the simple and hard samples, the classification model is batch-trained using the cross-entropy loss function, the selection optimization strategy loss function, and the neighbor aggregation strategy loss function to obtain the final general domain adaptive image classification model.
[0009] This invention provides a system for implementing a general domain adaptive image classification model, comprising:
[0010] The data module is used to set the source domain image dataset with classification labels and the target domain image dataset with different distributions and no classification labels as image data and classification labels for training the classification model;
[0011] The training module is used to train the closed-set classifier and the open-set classifier in the classification model based on the source domain image dataset and the cross-entropy loss function. The closed-set classifier outputs the classification probability of a sample belonging to each category in the source domain, and the multiple sub-binary classifiers of the open-set classifier output the within-class and out-class classification probabilities of a sample belonging to the corresponding category in the source domain and the maximum open-set entropy of the sample, respectively. The maximum open-set entropy is used to determine the confidence level of the sample.
[0012] The separation module is used to determine the most confusing category of the target domain image dataset samples based on the maximum open set entropy, set a transition zone in the most confusing category, select an optimization strategy based on the maximum open set entropy to separate simple samples outside the transition zone with confidence levels higher than a preset first threshold, and combine a neighbor aggregation strategy to separate difficult samples within the transition zone with confidence levels lower than a preset second threshold.
[0013] The joint module is used to batch train the classification model based on the simple and hard samples, by combining the cross-entropy loss function, the selection optimization strategy loss function, and the neighbor aggregation strategy loss function, to obtain the final general domain adaptive image classification model.
[0014] This invention also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the above-described general domain adaptive image classification model implementation method.
[0015] This invention also provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, implements the steps of the above-described general domain adaptive image classification model implementation method.
[0016] The following beneficial effects can be achieved by adopting the embodiments of the present invention: The embodiments of the present invention propose a new method for implementing a general domain adaptive image classification model to address the problems of domain difference and category difference in general domain adaptive scenarios. The embodiments of the present invention use a more robust criterion to measure the uncertainty of the target sample, and adaptively separate shared samples and "unknown" samples by the uncertainty of the target sample, while achieving weak domain alignment on the shared class. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a flowchart of the implementation method of the general domain adaptive image classification model according to an embodiment of the present invention;
[0019] Figure 2 This is a model framework diagram of the general domain adaptive image classification method based on maximum open set entropy optimization according to an embodiment of the present invention;
[0020] Figure 3 This is a schematic diagram of the implementation system of the general domain adaptive image classification model according to an embodiment of the present invention. Detailed Implementation
[0021] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.
[0022] Method Implementation Examples
[0023] According to embodiments of the present invention, a method for implementing a general domain adaptive image classification model is provided. Figure 1 This is a flowchart of the implementation method of the general domain adaptive image classification model according to an embodiment of the present invention, as follows: Figure 1 As shown, the general domain adaptive image classification model implementation method according to an embodiment of the present invention specifically includes:
[0024] Step S101: Set the source domain image dataset with classification labels and the target domain image dataset with different distributions and no classification labels as the image data and classification labels for training the classification model;
[0025] Step S102: Based on the source domain image dataset, train the closed-set classifier and the open-set classifier in the classification model using the cross-entropy loss function. The closed-set classifier outputs the classification probability of a sample belonging to each category in the source domain. The multiple sub-binary classifiers of the open-set classifier output the within-class and out-class classification probabilities of a sample belonging to the corresponding category in the source domain, as well as the maximum open-set entropy of the sample. The maximum open-set entropy is used to determine the confidence level of the sample.
[0026] The maximum open set entropy specifically includes:
[0027]
[0028] Among them, |L s | represents the number of categories in the source domain dataset s. This represents the i-th image data in the target domain t. and Let represent the in-class and out-class classification probabilities of the k-th open set subclass binary classifier, respectively, and j represent the error category that the model is most likely to confuse in the open set space;
[0029] Step S103: Determine the most confusing category of the target domain image dataset samples based on the maximum open set entropy, and set a transition zone in the most confusing category. Select an optimization strategy based on the maximum open set entropy to separate simple samples outside the transition zone with confidence levels higher than a preset first threshold, and combine the neighbor aggregation strategy to separate difficult samples inside the transition zone with confidence levels lower than a preset second threshold.
[0030] Step S104: Based on the simple and hard samples, the classification model is batch trained by combining the cross-entropy loss function, the selection optimization strategy loss function, and the neighbor aggregation strategy loss function to obtain the final general domain adaptive image classification model.
[0031] The cross-entropy loss function specifically includes:
[0032]
[0033] in, This represents the i-th image data in the source domain s and its corresponding classification label; l ce p is the standard cross-entropy loss function;c This represents the classification probability output by the closed set; and Let represent the in-class and out-class classification probabilities of the k-th open set subclass binary classifier, respectively, and j represent the error category that the model is most likely to confuse in the open set space;
[0034] The maximum open set output entropy selection optimization loss function specifically includes:
[0035]
[0036] Where m is a hyperparameter controlling the range of the transition region. This represents the maximum open set entropy of the i-th image data in the target domain t. It is the category corresponding to the maximum open set entropy of the i-th image data in the target domain t;
[0037] The neighbor aggregation loss function specifically includes:
[0038]
[0039]
[0040] in, The set of feature representations of all possible neighbors of the target sample, including all other target domain samples. and all source domain prototypes N j Represent the characteristics of the j-th neighbor. This indicates its transpose. s represents the feature representation of the i-th image data in the target domain t. i，j Represents target domain samples The similarity between the sample and its j-th neighbor; the temperature coefficient τ controls the compactness of the similarity distribution, implicitly affecting the number of neighbors for each sample; |L s | represents the number of categories in the source domain dataset s, N t This represents the number of images in the target domain t dataset;
[0041] The general domain adaptive image classification model specifically includes:
[0042]
[0043] in, This represents the i-th image data in the source domain s and its corresponding classification label. This represents the i-th image data in the target domain t. and These represent the source domain dataset and the target domain dataset, respectively. Represents the cross-entropy loss function. This indicates that the maximum open set output entropy is used to select the optimal loss function. This represents the neighbor aggregation loss function, where α and β are weight parameters used to balance the maximum open set output entropy selection optimization loss and the neighbor aggregation loss.
[0044] The general domain adaptive image classification method further includes:
[0045] The feature representation of image samples is extracted and stored through the feature extractor and memory module in the general domain adaptive image classification model;
[0046] The closed-set classifier and open-set classifier modules in the general domain adaptive image classification model predict the classification probability of image samples in the closed-set space and the open-set space.
[0047] The following describes the model framework of the general domain adaptive image classification method based on maximum open set entropy optimization, as described in the embodiments of the present invention. Figure 2 As shown, the above technical solutions of the embodiments of the present invention will be described in detail. The implementation method of the general domain adaptive image classification model of the embodiments of the present invention specifically includes the following steps:
[0048] Step 1) Prepare a source domain image dataset with classification labels and a target domain image dataset with a different distribution than the source domain data and without classification labels, as the image data and labels for model training;
[0049] Step 2) Supervised training of a closed-set classifier and an open-set classifier on the source domain dataset. The closed-set classifier predicts the probability of a sample belonging to each category in the source domain. The open-set classifier consists of multiple sub-binary classifiers. Each sub-binary classifier outputs the within-class and out-class probability of a sample belonging to the corresponding category in the source domain. The maximum open-set entropy output by the open-set classifier is used to measure the uncertainty of the sample. The larger the entropy, the higher the uncertainty of the sample. The smaller the entropy, the higher the confidence of the sample.
[0050] Step 3) Determine the most confusing class of the sample based on the maximum open set output entropy, and set a transition zone within this class. Use the maximum open set output entropy selection optimization strategy to separate simple samples with high confidence outside the transition zone, and combine the neighbor aggregation strategy to separate difficult samples with high uncertainty within the transition zone.
[0051] Step 4) The classification model is batch trained using the cross-entropy loss, maximum open set output entropy selection optimization loss, and neighbor aggregation loss, which are jointly supervised training methods, to obtain the final general domain adaptive image classification model.
[0052] The data read in this embodiment of the invention is a matrix composed of images after pixel value quantization, which facilitates the model's subsequent recognition of the visual features of the images.
[0053] In this embodiment of the invention, the cross-entropy loss function for supervised training of the closed-set classifier and the open-set classifier is:
[0054]
[0055] in, This represents the i-th image data in the source domain and its corresponding ground truth label; l ce p is the standard cross-entropy loss function; c This represents the output probability of closed set classification; and Let represent the in-class and out-class predicted probabilities of the k-th class open set sub-binary classifier, respectively, and j represent the error class that the model is most likely to confuse in the open set space.
[0056] In this embodiment of the invention, the maximum open set entropy output by the open set classifier is:
[0057]
[0058] Among them, |L s | indicates the number of categories in the source domain dataset; This represents the i-th image data in the target domain t.
[0059] In this embodiment of the invention, the optimization loss for selecting the maximum open set output entropy is:
[0060]
[0061] Where m is a hyperparameter that controls the range of the transition region; It is the category corresponding to the output entropy of the largest open set.
[0062] In this embodiment of the invention, the neighbor aggregation loss of the target sample is:
[0063]
[0064]
[0065] in, Represent the features of all possible neighbors of the target sample, including all other target samples. and all source domain prototypes s i，j Indicates target sample The similarity between the sample and its j-th neighbor; the temperature coefficient τ controls the compactness of the similarity distribution, implicitly affecting the number of neighbors for each sample; N t This indicates the number of images in the target domain dataset.
[0066] In this embodiment of the invention, the final image classification model is:
[0067]
[0068] in, and α and β represent the source domain dataset and the target domain dataset, respectively; α and β are weight parameters used to balance the maximum open set output entropy selection optimization loss and the neighbor aggregation loss.
[0069] This invention utilizes a closed-set classifier and an open-set classifier to predict the classification probability of samples. A novel maximum open-set entropy metric is proposed to measure sample uncertainty. Then, the most easily confused class is determined based on the maximum open-set output entropy, and a transition zone is defined within this class to promote sample separation. The overall classification model is batch-trained using a combination of supervised training cross-entropy loss, maximum open-set entropy optimization loss, and sample neighbor aggregation loss to obtain the final image classification model. The final general-domain adaptive image classification model of this invention also includes the following modules:
[0070] A feature extractor and memory module are used to acquire and store feature representations of image samples;
[0071] The closed-set classifier and open-set classifier modules are used to predict the classification probability of image samples in closed-set space and open-set space.
[0072] The method proposed in this invention can effectively solve the problems of domain offset and negative transfer of private samples in the target domain, and can improve the generalization ability of the model in different domains.
[0073] System Implementation Examples
[0074] According to embodiments of the present invention, a system for implementing a general domain adaptive image classification model is provided. Figure 3 This is a schematic diagram of the implementation system of the general domain adaptive image classification model according to an embodiment of the present invention, such as... Figure 3 As shown, the general domain adaptive image classification model implementation system according to an embodiment of the present invention specifically includes:
[0075] Data module 30 is used to set the source domain image dataset with classification labels and the target domain image dataset with different distribution and no classification labels as image data and classification labels for training the classification model;
[0076] Training module 32 is used to train the closed-set classifier and open-set classifier in the classification model based on the source domain image dataset and the cross-entropy loss function. The closed-set classifier outputs the classification probability of a sample belonging to each category of the source domain, and the multiple sub-binary classifiers of the open-set classifier output the intra-class and out-class classification probabilities of a sample belonging to the corresponding category of the source domain and the maximum open-set entropy of the sample, respectively. The maximum open-set entropy is used to determine the confidence level of the sample.
[0077] The separation module 34 is used to determine the most confusing category of the target domain image dataset samples based on the maximum open set entropy, set a transition zone in the most confusing category, select an optimization strategy based on the maximum open set entropy to separate simple samples outside the transition zone with confidence levels higher than a preset first threshold, and combine a neighbor aggregation strategy to separate difficult samples within the transition zone with confidence levels lower than a preset second threshold.
[0078] The joint module 36 is used to batch train the classification model based on the simple and hard samples, by combining the cross-entropy loss function, the selection optimization strategy loss function, and the neighbor aggregation strategy loss function, to obtain the final general domain adaptive image classification model.
[0079] Specifically, the maximum open set entropy is used for:
[0080]
[0081] Among them, |L s | represents the number of categories in the source domain dataset s. This represents the i-th image data in the target domain t. and Let represent the in-class and out-class classification probabilities of the k-th open set subclass binary classifier, respectively, and j represent the error category that the model is most likely to confuse in the open set space.
[0082] The cross-entropy loss function is specifically used for:
[0083]
[0084] in, This represents the i-th image data in the source domain s and its corresponding classification label; l ce p is the standard cross-entropy loss function; c This represents the classification probability output by the closed set; and Let represent the in-class and out-class classification probabilities of the k-th open set subclass binary classifier, respectively, and j represent the error category that the model is most likely to confuse in the open set space;
[0085] The maximum open set output entropy selection optimization loss function is specifically used for:
[0086]
[0087] Where m is a hyperparameter controlling the range of the transition region. This represents the maximum open set entropy of the i-th image data in the target domain t. It is the category corresponding to the maximum open set entropy of the i-th image data in the target domain t;
[0088] The neighbor aggregation loss function is specifically used for:
[0089]
[0090]
[0091] in, The set of feature representations of all possible neighbors of the target sample, including all other target domain samples. and all source domain prototypes N j Represent the characteristics of the j-th neighbor. This indicates its transpose. s represents the feature representation of the i-th image data in the target domain t. i，j Represents target domain samples The similarity between the sample and its j-th neighbor; the temperature coefficient τ controls the compactness of the similarity distribution, implicitly affecting the number of neighbors for each sample; |L s | represents the number of categories in the source domain dataset s, N t This represents the number of images in the target domain t dataset.
[0092] The general domain adaptive image classification model is specifically used for:
[0093]
[0094] in, This represents the i-th image data in the source domain s and its corresponding classification label. This represents the i-th image data in the target domain t. and These represent the source domain dataset and the target domain dataset, respectively. Represents the cross-entropy loss function. This indicates that the maximum open set output entropy is used to select the optimal loss function. Let represent the neighbor aggregation loss function, where α and β are weight parameters used to balance the maximum open set output entropy selection optimization loss and the neighbor aggregation loss.
[0095] The system further includes:
[0096] The classification module is used to extract and store the feature representation of image samples through the feature extractor and memory module in the general domain adaptive image classification model; and to predict the classification probability of image samples in the closed set space and the open set space through the closed set classifier and open set classifier modules in the general domain adaptive image classification model.
[0097] The following is in conjunction with the appendix Figure 2 The technical solutions described above in the embodiments of the present invention will be explained in detail.
[0098] Step 1) Prepare a source domain image dataset with classification labels and a target domain image dataset with a different distribution than the source domain data and without classification labels, as the image data and labels for model training;
[0099] Step 2) Supervised training of a closed-set classifier and an open-set classifier is performed on the source domain dataset. The closed-set classifier predicts the probability that a sample belongs to each category in the source domain. The open-set classifier consists of multiple sub-classifiers, each outputting the within-class and out-class probabilities of a sample belonging to the corresponding category in the source domain. The maximum open-set entropy output by the open-set classifier is used to measure the uncertainty of the sample; a higher entropy reflects higher sample uncertainty, while a lower entropy reflects higher sample confidence.
[0100] Step 3) Determine the most confusing class of the sample based on the maximum open set output entropy, and set a transition zone within this class. Use the maximum open set output entropy selection optimization strategy to separate simple samples with high confidence outside the transition zone, and combine the neighbor aggregation strategy to separate difficult samples with high uncertainty within the transition zone.
[0101] Step 4) The classification model is batch trained using the cross-entropy loss, maximum open set output entropy selection optimization loss, and neighbor aggregation loss, which are jointly supervised training methods, to obtain the final general domain adaptive image classification model.
[0102] Wherein, the cross-entropy loss function of the supervised training closed-set classifier and open-set classifier in step 2) is:
[0103]
[0104] in, This represents the i-th image data in the source domain and its corresponding ground truth label; l ce p is the standard cross-entropy loss function; c This represents the output probability of closed set classification; and Let represent the in-class and out-of-class predicted probabilities of the k-th open set sub-binary classifier, respectively, and j represent the error class most easily confused by the model in the open set space. The maximum open set entropy output by the open set classifier is:
[0105]
[0106] Among them, |L s | indicates the number of categories in the source domain dataset; This represents the i-th image data in the target domain.
[0107] Step 3) The optimization loss for selecting the maximum open set output entropy is:
[0108]
[0109] Where m is a hyperparameter that controls the range of the transition region; It is the category corresponding to the output entropy of the largest open set.
[0110] The neighbor aggregation loss is:
[0111]
[0112]
[0113] in, Represent the features of all possible neighbors of the target sample, including all other target samples. and all source domain prototypes s i，j Indicates target sample The similarity between the sample and its j-th neighbor; the temperature coefficient τ controls the compactness of the similarity distribution, implicitly affecting the number of neighbors for each sample; N t This indicates the number of images in the target domain dataset.
[0114] Step 4) The joint supervised training loss for image classification is:
[0115]
[0116] in, and Let represent the source domain dataset and the target domain dataset, respectively; α and β are weight parameters used to balance the maximum open set output entropy selection optimization loss and the neighbor aggregation loss. The final general domain adaptive image classification model includes the following modules:
[0117] A feature extractor and memory module are used to acquire and store feature representations of image samples;
[0118] The closed-set classifier and open-set classifier modules are used to predict the classification probability of samples in the closed-set space and the open-set space.
[0119] In summary, the training algorithm based on maximum open set entropy optimization proposed in this embodiment of the invention can effectively solve the problems of domain offset and negative transfer of private samples in the target domain, and can improve the generalization ability of the model in different domains.
[0120] Device Example 1
[0121] This invention provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, performs the steps described in the method embodiment.
[0122] Device Example 2
[0123] This invention provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, performs the steps described in the method embodiment.
[0124] The computer-readable storage media described in this embodiment include, but are not limited to, ROM, RAM, disk, or optical disk.
[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for implementing a general domain adaptive image classification model, characterized in that, include: Set the source domain image dataset with classification labels and the target domain image dataset with a different distribution than the source domain data and no classification labels as the image data and classification labels for training the classification model; Based on the source domain image dataset, the closed-set classifier and open-set classifier in the classification model are trained using the cross-entropy loss function. The closed-set classifier outputs the classification probability of a sample belonging to each category in the source domain. Multiple sub-binary classifiers of the open-set classifier output the within-class and out-class classification probabilities of a sample belonging to the corresponding category in the source domain, as well as the maximum open-set entropy of the sample. The maximum open-set entropy is used to determine the confidence level of the sample; specifically, the maximum open-set entropy includes: Official 1; in, Represents the source domain The number of categories in the dataset. Represents the target domain The Middle Image data, and Let represent the in-class and out-class classification probabilities of the R-th open set sub-binary classifier, respectively, where R represents the error class that the model is most likely to confuse in the open set space; The most easily confused category of the target domain image dataset samples is determined based on the maximum open set entropy, and a transition zone is set in the most easily confused category. An optimization strategy is selected based on the maximum open set entropy to separate simple samples outside the transition zone with confidence levels higher than a preset first threshold, and a neighbor aggregation strategy is combined to separate difficult samples within the transition zone with confidence levels lower than a preset second threshold. Based on the simple and hard samples, the classification model is batch-trained using the cross-entropy loss function, the selection optimization strategy loss function, and the neighbor aggregation strategy loss function to obtain the final general domain adaptive image classification model.
2. The method according to claim 1, characterized in that, The cross-entropy loss function specifically includes: Official 2; in, Represents the source domain The Middle Each image data point and its corresponding classification label; The standard cross-entropy loss function; This represents the classification probability output by the closed set; Indicates the first The in-class classification probability of the open set subclass binary classifier, where R represents the error class that the model is most likely to confuse in the open set space; The loss function for the maximum open set output entropy selection optimization strategy specifically includes: Official 3; in, It is a hyperparameter that controls the range of the transition region. Represents the target domain The Middle Maximum open set entropy of image data The target domain The Middle The category corresponding to the maximum open set entropy of each image data; The loss function of the neighbor aggregation strategy specifically includes: Official 4; Official 5; in, The set of feature representations of all possible neighbors of the target sample, including all other target domain samples. and all source domain prototypes , Indicates the first Characteristics of a neighbor, This indicates its transpose. Represents the target domain The Middle Feature representation of image data Represents target domain samples Its first Similarity between neighbors; temperature coefficient It controls the compactness of the similarity distribution and implicitly affects the number of neighbors for each sample; Represents the source domain The number of categories in the dataset. Represents the target domain The number of images in the dataset.
3. The method according to claim 1, characterized in that, The general domain adaptive image classification model specifically includes: Official 6; in, Represents the source domain The Middle Each image data point and its corresponding classification label, Represents the target domain The Middle Image data, and These represent the source domain dataset and the target domain dataset, respectively. Represents the cross-entropy loss function. This represents the loss function for selecting the optimal strategy based on the maximum open set output entropy. Represents the neighbor aggregation loss function. and It is a weight parameter used to balance the maximum open set output entropy selection optimization loss and the neighbor aggregation loss.
4. The method according to claim 1, characterized in that, The general domain adaptive image classification model implementation method further includes: The feature representation of image samples is extracted and stored through the feature extractor and memory module in the general domain adaptive image classification model; The closed-set classifier and open-set classifier modules in the general domain adaptive image classification model predict the classification probability of image samples in the closed-set space and the open-set space.
5. A system for implementing a general domain adaptive image classification model, characterized in that, include: The data module is used to set the source domain image dataset with classification labels and the target domain image dataset with different distributions and no classification labels as image data and classification labels for training the classification model; The training module is used to train the closed-set classifier and open-set classifier in the classification model based on the source domain image dataset and the cross-entropy loss function. The closed-set classifier outputs the classification probability of a sample belonging to each category in the source domain, and the multiple sub-binary classifiers of the open-set classifier output the within-class and out-class classification probabilities of a sample belonging to the corresponding category in the source domain, as well as the maximum open-set entropy of the sample. The maximum open-set entropy is used to determine the confidence level of the sample; specifically, the maximum open-set entropy includes: Official 1; in, Represents the source domain The number of categories in the dataset. Represents the target domain The Middle Image data, and Let represent the in-class and out-class classification probabilities of the R-th open set sub-binary classifier, respectively, where R represents the error class that the model is most likely to confuse in the open set space; The separation module is used to determine the most confusing category of the target domain image dataset samples based on the maximum open set entropy, set a transition zone in the most confusing category, select an optimization strategy based on the maximum open set entropy to separate simple samples outside the transition zone with confidence levels higher than a preset first threshold, and combine a neighbor aggregation strategy to separate difficult samples within the transition zone with confidence levels lower than a preset second threshold. The joint module is used to batch train the classification model based on the simple and hard samples, by combining the cross-entropy loss function, the selection optimization strategy loss function, and the neighbor aggregation strategy loss function, to obtain the final general domain adaptive image classification model.
6. The system according to claim 5, characterized in that, The cross-entropy loss function is specifically used for: Official 2; in, Represents the source domain The Middle Each image data point and its corresponding classification label; The standard cross-entropy loss function; This represents the classification probability output by the closed set; Indicates the first The in-class classification probability of the open set subclass binary classifier, where R represents the error class that the model is most likely to confuse in the open set space; The loss function for the maximum open set output entropy selection optimization strategy is specifically used for: Official 3; in, It is a hyperparameter that controls the range of the transition region. Represents the target domain The Middle Maximum open set entropy of image data The target domain The Middle The category corresponding to the maximum open set entropy of each image data; The neighbor aggregation strategy loss function is specifically used for: Official 4; Official 5; in, The set of feature representations of all possible neighbors of the target sample, including all other target domain samples. and all source domain prototypes , Indicates the first Characteristics of neighbors This indicates its transpose. Represents the target domain The Middle Feature representation of image data Represents target domain samples Its first Similarity between neighbors; temperature coefficient It controls the compactness of the similarity distribution and implicitly affects the number of neighbors for each sample; Represents the source domain The number of categories in the dataset. Represents the target domain The number of images in the dataset; The general domain adaptive image classification model is specifically used for: Official 6; in, Represents the source domain The Middle Each image data point and its corresponding classification label, Represents the target domain The Middle Image data, and These represent the source domain dataset and the target domain dataset, respectively. Represents the cross-entropy loss function. This represents the loss function for selecting the optimal strategy based on the maximum open set output entropy. Represents the neighbor aggregation loss function. and It is a weight parameter used to balance the maximum open set output entropy selection optimization loss and the neighbor aggregation loss.
7. The system according to claim 5, characterized in that, The system further includes: The classification module is used to extract and store the feature representation of image samples through the feature extractor and memory module in the general domain adaptive image classification model; and to predict the classification probability of image samples in the closed set space and the open set space through the closed set classifier and open set classifier modules in the general domain adaptive image classification model.
8. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the general domain adaptive image classification model implementation method as described in any one of claims 1 to 4.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an implementation program for information transmission, which, when executed by a processor, implements the steps of the general domain adaptive image classification model implementation method as described in any one of claims 1 to 4.