A transfer learning method based on correlation of global features and intrinsic features

Abstract

A kind of transfer learning method based on global feature and internal feature association, comprising extracting the depth feature of sample by deep neural network;From target domain, select a certain proportion of sample as labeled target domain sample, the rest sample belongs to unlabeled target domain sample;Adopt semi-supervised adversarial transfer learning method, learn the base model of target domain;According to depth feature, respectively to the whole sample of source domain, target domain unlabeled sample respectively carries out unsupervised clustering, each clustering is called a component;The component of source domain and target domain is matched, and each matching is called a pair;For each pair, fine-tuning is carried out based on base model, so that it is more in line with the target domain component in the pair.Meanwhile, the overall distribution of features and the internal correlation between features are considered, which can better realize knowledge transfer, thereby better solving the problems of time-consuming and labor-intensive artificial labeling in object recognition and other fields, and improving the transfer efficiency.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence transfer learning technology, and in particular to a transfer learning method based on the correlation between global features and intrinsic features. Background Technology

[0002] Deep learning has achieved great success in image classification. However, this success relies on models learning a large number of parameters and having a large amount of labeled data to support them. But when faced with new scenarios (or slightly altered tasks), the previously learned models often cannot be directly applied, requiring the training of new models for the new scenarios. In practical classification problems, it is difficult to obtain the large-scale data needed to build models, and manually labeling data is actually very inefficient. Models trained on large amounts of manually labeled data rarely achieve ideal results on real-world data. Therefore, transfer learning is used to address these problems.

[0003] Existing image classification transfer methods are based on global category information and do not consider the similarities and differences of the intrinsic subtle features of the samples. This greatly reduces the performance of the model and leads to the problem of low transfer efficiency. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of the prior art by proposing a transfer learning method based on the correlation between global features and intrinsic features. This technical solution is applied to the field of image classification learning. First, deep features of samples are extracted using a deep neural network. A certain proportion of samples are selected from the target domain and these samples are designated as labeled target domain samples, while the remaining samples are unlabeled target domain samples. A semi-supervised adversarial transfer learning method is used to learn the base model of the target domain. Based on the deep features, unsupervised clustering is performed on all samples in the source domain and the unlabeled samples in the target domain, thereby grouping samples with similar intrinsic features together. Each cluster is called a component. The components of the source domain and the target domain are matched, and each match is called a pair. For each pair, the base model is fine-tuned to better match the target domain components in that pair. Simultaneously considering the overall distribution of features and the intrinsic correlation between features, this method can better achieve knowledge transfer, thus better solving the problems of time-consuming and labor-intensive manual annotation in fields such as object recognition, and improving transfer efficiency.

[0005] To address the above problems, this invention proposes a transfer learning method based on the correlation between global features and intrinsic features, comprising the following steps:

[0006] Step A: Prepare samples by dividing the labeled image data of the current scene into the source domain and the unlabeled image data of the new scene into the target domain.

[0007] Step B: Extract depth features by using a deep neural network to extract the depth features of the samples;

[0008] Step C: Sample selection. Randomly select a preset number of samples from the target domain and obtain the labels of the selected samples.

[0009] Step D: Train the base model. Based on the source domain in Step A, the labeled samples selected in Step C, and the unlabeled samples not selected in Step C, train the base model for the target domain.

[0010] Step E: Extract components. Perform unsupervised clustering on the source domain and the unlabeled samples that were not selected in Step C. Each cluster obtained is a component.

[0011] Step F: Component matching, matching the components in the target domain with the corresponding source domain components, so that each component in the target domain forms a pair with its corresponding source domain component;

[0012] Step G: Fine-tune the base models, performing fine-tuning based on each paired base model.

[0013] Preferably, step C includes labeling the source domain as D. S The target domain is labeled as D. T ;

[0014] The target domain D T Divided into D T ={D TL ∪D TU}, where D TL D represents a randomly selected sample, where the labels of the randomly selected sample are known. TU This indicates samples that were not selected; the label for unselected samples is unknown.

[0015] The D TL The sample size is much smaller than D TU The number of samples.

[0016] Preferably, step D includes a source domain D with labels. S Randomly selected sample D TL and unselected sample D TU For the target domain D T Training the base model;

[0017] Training the base model includes two overlapping steps: adversarial generation and adversarial training. In the adversarial generation step, a concatenation sample is generated for each original sample; this concatenation sample is used to fill the source domain D. S and target domain D TThe feature distribution differences; classifier C and discriminator D are trained based on adversarial training steps.

[0018] Preferably, the adversarial generation step is performed according to Formula 1 and Formula 2;

[0019]

[0020]

[0021] in:

[0022] Represents the connection sample of the source domain;

[0023] Represents the connecting samples of the target domain;

[0024] w represents the number of iterations for the adaptive sample. When w = 0, This represents the original sample corresponding to the source domain. This represents the original sample corresponding to the target domain;

[0025] α and β represent the equilibrium parameters of Equation 1 and Equation 2, respectively;

[0026] θ C These represent the model parameters of classifier C;

[0027] θ D This represents the model parameters of the discriminator;

[0028] Let C represent the loss function of the classifier C;

[0029] This represents the loss function of the discriminator D;

[0030] Preferably, in the adversarial generation step, Euclidean distance is used to constrain the feature distribution of the connecting samples to lie between the source and target domains.

[0031] In Formula 1, Represents the Euclidean distance-based constraints between the original samples and the connecting samples in the source domain;

[0032] In Formula 2, This represents the Euclidean distance-based constraint between the original samples and the connecting samples in the target domain.

[0033] Preferably, the adversarial training steps are performed according to Formulas 3 and 4;

[0034]

[0035]

[0036] in:

[0037] y represents the real label;

[0038] x represents the original sample;

[0039] x * Indicates connecting samples;

[0040] C represents the classifier;

[0041] D S This represents the original sample from the source domain with labels.

[0042] D T This represents the original sample of the target domain with labels;

[0043] This represents a sample connecting the source and target domains.

[0044] Represents cross-entropy;

[0045] θ C These represent the model parameters of classifier C;

[0046] Let C represent the loss function of the classifier C;

[0047] Represents the expected value of the function;

[0048] adv indicates adversarial training;

[0049] The classification loss function is related to the connecting samples.

[0050] Preferably, it also includes using a classifier C and a discriminator D to obtain labeled original samples D. S and D TL The loss;

[0051] The loss of classifier C is calculated from the original sample x and the adaptive sample x. * composition;

[0052] This includes using discriminator D to obtain unlabeled raw samples D. TU The loss is calculated using formulas five and six.

[0053]

[0054]

[0055] in:

[0056] This represents the loss function of the discriminator D;

[0057] θ d These represent the model parameters of the discriminator D;

[0058] D represents the discriminator D;

[0059] D ST This represents all original samples containing both the source and target domains;

[0060] This represents the concatenated sample corresponding to all original samples containing both the source and target domains.

[0061] Represents the expectation of the function;

[0062] x S This represents a single original sample belonging to the source domain;

[0063] x T This represents a single original sample belonging to the target domain;

[0064] This represents a single connected sample belonging to the source domain;

[0065] This represents a single connection sample belonging to the target domain;

[0066] adv indicates adversarial training;

[0067] This represents the discriminant loss function associated with the connecting samples.

[0068] Preferably, in step E, the extraction component includes processing the labeled source domain D based on the depth features of the sample. S And the unselected sample D TU Perform unsupervised clustering;

[0069] After clustering, samples within the same component contain at least one label.

[0070] Preferably, step F includes component matching using a bipartite graph, with MMD distance used to measure the distance between two components.

[0071] Preferably, source domain D S and target domain D T The components form two subsets in the bipartite graph, and the two subsets do not overlap.

[0072] Get target domain D T Any component in the source domain D S The MMD distance of all components is used as the weight of the corresponding edge in the bipartite graph.

[0073] For the target domain D T For each component, select the edge with the smallest weight (i.e., the smallest MMD distance) to match the corresponding source domain D. S Components.

[0074] The beneficial effects of the technical solution of this invention are:

[0075] This invention first extracts deep features, then aligns the data distributions of the source and target domains as much as possible from a global perspective to learn a base model for the target domain. Next, it fine-tunes the base model based on the inherent relationships between features, thus considering the intrinsic connections between features. It employs an unsupervised approach to extract components from the source and target domains separately. Then, it uses bipartite graphs for component matching and MMD (Multi-Dimensional Mapping) to measure the distance between different components, allowing for pairing of the closest components from different domains. This enables more targeted fine-tuning of the base model. Simultaneously considering the overall distribution of features and the intrinsic relationships between features, it achieves better knowledge transfer, thus better solving the problems of time-consuming and labor-intensive manual annotation in fields such as object recognition, and improving transfer efficiency. Attached Figure Description

[0076] Figure 1 This is a flowchart of a transfer learning method based on the association between global features and intrinsic features according to an embodiment of the present invention;

[0077] Figure 2 This is a schematic diagram of the framework of a transfer learning method based on the association between global features and intrinsic features, according to an embodiment of the present invention. Detailed Implementation

[0078] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0079] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0080] The terms "comprising" and "having," and any variations thereof, in the specification and claims of this invention are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or ends.

[0081] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0082] Deep learning has achieved great success in image classification. However, this success relies on models learning a large number of parameters and having a large amount of labeled data to support them. But when faced with new scenarios (or slightly different tasks), the previously learned models often cannot be directly applied, requiring the training of new models for the new scenarios. In practical classification problems, it is difficult to obtain the large-scale data needed to build models, and manually labeled data is actually very inefficient. Models trained on large amounts of manually labeled data rarely achieve ideal results on real-world data. Therefore, this application addresses this problem by utilizing transfer learning.

[0083] Existing image classification transfer learning methods are all based on global category information, neglecting the similarities and differences in the intrinsic subtle features of samples. This significantly reduces model performance. In other words, existing techniques align the data distributions of the source and target domains from the perspective of global features to learn a classification model for the target domain. These techniques do not consider the intrinsic relationships between features or the influence of intrinsic features on samples, making it difficult to learn an ideal target domain model. To address these issues, this invention proposes a transfer learning method based on the correlation between global features and intrinsic features, such as... Figure 1 and Figure 2 As shown, it includes the following steps:

[0084] Step A: Prepare samples by dividing the labeled image data of the current scene into the source domain and the unlabeled image data of the new scene into the target domain.

[0085] In this embodiment, the sample includes a source domain and a target domain, and the source domain is represented by D. S The target domain is represented by D. T In this embodiment, the source domain can be understood as old data, and the target domain can be understood as new data. Specifically, in image classification transfer, the labeled image data of the current scene can be referred to as the source domain, and the unlabeled image data of the new scene can be referred to as the target domain.

[0086] The feature distributions of the samples differ, therefore the corresponding models also differ. For example, photos of students in their final year of junior high and senior high school may differ, even though they are still the same students; the corresponding facial recognition program must be adjusted. If the target domain D is set manually... T Labeling all samples would be extremely labor-intensive; therefore, the technical solution of this application is to first label the target domain D. T A small portion of the samples were labeled, with the target domain D. T All samples are unlabeled, so labeling is required. These labeled samples will be denoted as D in the following text. TL Then combine with source domain D S For the target domain D T Training the classification model. In this embodiment, the above operation can be understood as applying the classification model of the current scene to the new scene using transfer learning.

[0087] Step B: Extract depth features by using a deep neural network to extract the depth features of the samples;

[0088] Original samples (referring to the initial samples) typically have high dimensionality. Often, only relatively low dimensionality is needed to reflect the main characteristics of the sample, thus reducing computational cost. The advantage of deep neural networks lies in extracting deep features, which can essentially reflect the main characteristics of the sample. In this embodiment, deep neural networks such as ResNet and VGG can be used to extract the deep features of the sample. This technology is existing and will not be elaborated upon in this embodiment. The sample after extracting deep features can still be called the original sample.

[0089] Step C: Sample selection. Randomly select a preset number of samples from the target domain and obtain the labels of the selected samples.

[0090] Specifically, step C includes labeling the source domain as D. S The target domain is labeled as D. T ;

[0091] The target domain D T Divided into D T ={D TL ∪D TU}, where D TL D represents a randomly selected sample, where the labels of the randomly selected sample are known. TU This indicates samples that were not selected; the label for unselected samples is unknown.

[0092] The D TL The sample size is much smaller than D TU The number of samples.

[0093] In this embodiment, the specific operation is to start from the target domain D. T 5% of the labeled samples were randomly selected and labeled as D. TL Target domain D T The remaining samples are labeled as target domain D. TU .

[0094] Furthermore, in this embodiment, the D mentioned above... TL The labels are known and D. TU The meaning of "unknown label" can be understood as the selected samples in the target domain DT being known to belong to a certain label (this label can be understood as, in the selection of D...). TL Then, D was manually adjusted. TL (Labeling), the remaining samples that were not selected belong to the unknown label.

[0095] Step D: Train the base model. Based on the source domain in Step A, the labeled samples selected in Step C, and the unlabeled samples not selected in Step C, train the base model for the target domain.

[0096] Preferably, step D includes a source domain D with labels. S Randomly selected sample D TL and unselected sample D TU For the target domain D T Training the base model; the base model in this embodiment is the classification model mentioned above, and will be consistently referred to as the base model in the following text;

[0097] Training the base model includes two overlapping steps: adversarial generation and adversarial training. In the adversarial generation step, a concatenation sample is generated for each original sample; this concatenation sample is used to fill the source domain D. S and target domain D T The feature distribution differences; classifier C and discriminator D are trained based on adversarial training steps.

[0098] In this embodiment, the adversarial generation step generates a connecting sample, which is also called a fake sample or adversarial sample. Each original sample will generate a connecting sample; while the adversarial training step is used to train the classifier C and the discriminator D.

[0099] In this embodiment, the connecting samples can fill the source domain D. S and target domain D T The feature distributions differ, and both classifier C and discriminator D are deep neural networks.

[0100] Preferably, the adversarial generation process is as shown in Formula 1 and Formula 2;

[0101]

[0102]

[0103] in:

[0104] Represents the connection sample of the source domain;

[0105] Represents the connecting samples of the target domain;

[0106] w represents the number of iterations for the adaptive sample. When w = 0, This represents the original sample corresponding to the source domain. This represents the original sample corresponding to the target domain;

[0107] α and β represent the equilibrium parameters of Equation 1 and Equation 2, respectively;

[0108] θ C These represent the model parameters of classifier C;

[0109] θ D These represent the model parameters of the discriminator D;

[0110] Let C represent the loss function of the classifier C;

[0111] This represents the loss function of the discriminator D;

[0112] To ensure that the feature distribution of the connecting samples lies between the source and target domains, this embodiment uses Euclidean distance for constraint.

[0113] In Formula 1, Represents the Euclidean distance-based constraints between the original samples and the connecting samples in the source domain;

[0114] In Formula 2, This represents the Euclidean distance-based constraint between the original samples and the connecting samples in the target domain.

[0115] Preferably, the adversarial training process is as shown in Formulas 3 and 4;

[0116]

[0117]

[0118] in:

[0119] y represents the real label;

[0120] x represents the original sample;

[0121] x *Indicates connecting samples;

[0122] C represents the classifier;

[0123] D S This represents the original sample with labels (i.e., the source domain);

[0124] This represents the concatenated samples corresponding to all original samples in the source and target domains.

[0125] Represents cross-entropy;

[0126] θ C These represent the model parameters of classifier C;

[0127] Let C represent the loss function of the classifier C;

[0128] Represents the expectation of the function;

[0129] adv indicates adversarial training;

[0130] The classification loss function is related to the connecting samples.

[0131] Preferably, it also includes using a classifier C and a discriminator D to obtain labeled original samples D. S and D TL The loss;

[0132] In this embodiment, for sample D with a label S and D TL The loss of classifier C and discriminator D is calculated using the classifier and discriminator respectively. Gradient information is used to obtain and iteratively update the connecting samples. For unlabeled samples D... TU The loss is calculated using only the discriminator D, while gradient information is used to generate connecting samples.

[0133] The loss of classifier C is calculated from the original sample x and the adaptive sample x. * composition;

[0134] In this embodiment, for sample D with a label S and D TL By calculating cross-entropy To calculate the loss, for unlabeled sample D TU The loss is calculated by calculating the distance output by the discriminator D;

[0135] This includes using discriminator D to obtain unlabeled raw samples D. TU The loss is calculated using formulas five and six.

[0136]

[0137]

[0138] in:

[0139] This represents the loss function of the discriminator D;

[0140] θ d These represent the model parameters of the discriminator D;

[0141] D represents the discriminator D;

[0142] D ST This represents all original samples containing both the source and target domains;

[0143] This represents the concatenated samples corresponding to all original samples in the source and target domains.

[0144] Represents the expectation of the function;

[0145] x S This represents a single original sample belonging to the source domain;

[0146] x T This represents a single original sample belonging to the target domain;

[0147] This represents a single connected sample belonging to the source domain;

[0148] This represents a single connection sample belonging to the target domain;

[0149] adv indicates adversarial training;

[0150] The classification loss function is related to the connecting samples.

[0151] Discriminator D is used to distinguish whether a sample comes from the source domain or the target domain. Although the connecting samples corresponding to the source domain have learned information from the target domain, and the connecting samples corresponding to the target domain have also learned information from the source domain, discriminator D can still correctly determine whether any connecting sample corresponds to the source domain or the target domain.

[0152] Step E: Extract components. Perform unsupervised clustering on the source domain and the unlabeled samples that were not selected in Step C. Each cluster obtained is a component.

[0153] Furthermore, in step E, the extraction component includes processing the labeled source domain D based on the depth features of the sample. S And the unselected sample D TUPerform unsupervised clustering;

[0154] Clustering considers only the features of the samples, not their labels. Samples within the same component share similar intrinsic features, and this association is unaffected by the sample category. Therefore, a single component often contains samples from multiple categories. The number of components depends on the actual situation of the samples. S and D T The number of components is not necessarily the same.

[0155] Furthermore, after clustering, samples within the same component contain at least one label.

[0156] In this embodiment, D obtained through step B S and D TU The deep feature extraction component combines abstract knowledge in the image, such as properties, complex contours, and semantic information, to perform unsupervised clustering. After clustering, the same component often contains samples with multiple labels, which means that although the labels are different, other information can be the same.

[0157] Step F: Component matching, matching the components in the target domain with the corresponding source domain components, so that each component in the target domain forms a pair with its corresponding source domain component;

[0158] Preferably, step F includes component matching using a bipartite graph, with MMD distance used to measure the distance between two components.

[0159] Preferably, source domain D S and target domain D T The components form two subsets in the bipartite graph, and the two subsets do not overlap.

[0160] Get target domain D T Any component in the source domain D S The MMD distance of all components is used as the weight of the corresponding edge in the bipartite graph.

[0161] For each component of the target domain DT, select the edge with the smallest weight (i.e., the edge with the smallest MMD distance) to match the corresponding source domain D. S Components.

[0162] In this embodiment, the ultimate goal of this technical solution is to train a suitable classifier for the target domain, that is, to obtain a suitable base model, so as to classify the image data of the new scene using the base model. Therefore, it is only necessary to find the corresponding pair for each component of the target domain.

[0163] This embodiment uses a bipartite graph for component matching, and uses MMD distance to measure the distance between two components, allowing different components in the target domain to be paired with the same component in the source domain.

[0164] Specifically, in a bipartite graph, two disjoint subsets are respectively represented by D. S and D T The components constitute D. T For any component, this technical solution calculates its relationship with D. S The MMD distance of all components is used as the weight of the corresponding edge in the bipartite graph. For D T For each component, this technical solution selects the edge with the smallest weight (smallest MMD distance) to match the corresponding D. S The components are then paired in this way.

[0165] Based on distance, D S and D T The components are matched. In each final pairing, D... S and D T They share some kind of implicit or similar intrinsic connection (this intrinsic connection is implicit, not necessarily describable, nor needing to be describable, although it objectively exists). For example, D. S Apple images and D T The basketball images in the images form a pair because they are all spheres in shape and share some kind of information.

[0166] Step G: Fine-tune the base models, performing fine-tuning based on each paired base model.

[0167] Base models align features between the source and target domains from a global perspective, but existing models do not consider the intrinsic relationships between features, making it difficult to achieve ideal results. This application addresses this issue using bipartite graphs. Bipartite graphs can find the most similar component from the source domain for each component in the target domain, resulting in better transfer performance between similar components. Each component in the target domain forms a pair with a corresponding component in the source domain; that is, the number of pairs equals the number of components in the target domain. Then, the base model is fine-tuned based on each pair. During fine-tuning, the training method is basically the same as the base model training method mentioned above, except that samples are no longer selected from the components in the target domain as labeled samples.

[0168] Compared to existing technologies, this application considers both the overall distribution of features and the intrinsic relationships between features. First, deep features are extracted. Then, the data distributions of the source and target domains are aligned as much as possible from a global perspective to learn the base model for the target domain. Next, the base model is fine-tuned based on the intrinsic relationships between features, thus taking into account the intrinsic relationships between features. Components of the source and target domains are extracted separately using an unsupervised approach. Then, component matching is performed based on a bipartite graph, and the distance between different components is measured based on MMD (Multi-Dimensional Mapping). This allows the closest components from different domains to be paired, enabling more targeted fine-tuning of the base model.

[0169] The technical principles of the present invention have been described above with reference to specific embodiments. These descriptions are merely for explaining the principles of the invention and should not be construed as limiting the scope of protection of the invention in any way. Based on this explanation, those skilled in the art can readily conceive of other specific embodiments of the invention without inventive effort, and these embodiments will all fall within the scope of protection of the present invention.

Claims

1. A transfer learning method based on the correlation between global features and intrinsic features, characterized in that: Includes the following steps: Step A: Prepare samples by dividing the labeled image data of the current scene into the source domain and the unlabeled image data of the new scene into the target domain. Step B: Extract depth features by using a deep neural network to extract the depth features of the samples; Step C: Sample selection. Randomly select a preset number of samples from the target domain and label the selected samples. Step C includes marking the source domain as The target domain is marked as ; Step D: Train the base model. Based on the source domain in Step A, the labeled samples selected in Step C, and the unlabeled samples not selected in Step C, train the base model for the target domain. Step E: Extract components. Perform unsupervised clustering on the source domain and the unlabeled samples that were not selected in Step C. Each cluster obtained is a component. Step F: Component matching, matching the components in the target domain with the corresponding source domain components, so that each component in the target domain forms a pair with its corresponding source domain component; Step F includes component matching using a bipartite graph, and measuring the distance between two components using MMD distance, including: source domain and target domain The components form two subsets in the bipartite graph, and the two subsets do not overlap. Get target domain Any component and source domain The MMD distance of all components is used as the weight of the corresponding edge in the bipartite graph. For the target domain For each component, select the edge with the smallest weight to match the corresponding source domain. Components; Step G: Fine-tune the base models, performing fine-tuning based on each paired base model.

2. The transfer learning method based on the correlation between global features and intrinsic features according to claim 1, characterized in that: The target domain Divided into ,in, This represents a randomly selected sample, whose labels are known. This indicates samples that were not selected; the label for unselected samples is unknown. The The number of samples is much smaller than The number of samples.

3. The transfer learning method based on the correlation between global features and intrinsic features according to claim 2, characterized in that: Step D includes a source domain based on the label. Randomly selected samples and unselected samples To be the target domain Training the base model; Training the base model includes two overlapping steps: adversarial generation and adversarial training. In the adversarial generation step, a concatenation sample is generated for each original sample, and the concatenation sample is used to fill the source domain. and target domain The characteristic distribution differences; Classifier C and discriminator D are trained based on adversarial training steps.

4. The transfer learning method based on the correlation between global features and intrinsic features according to claim 3, characterized in that: Perform the adversarial generation steps according to Formula 1 and Formula 2; --Formula 1; --Formula 2; in: Represents the connection sample of the source domain; Represents the connecting samples of the target domain; w represents the number of iterations for the adaptive sample. When w=0, This represents the original sample corresponding to the source domain. This represents the original sample corresponding to the target domain; and These represent the balance parameters of Formula 1 and Formula 2, respectively; These represent the model parameters of classifier C; This represents the model parameters of the discriminator; Let C represent the loss function of the classifier C; Let represent the loss function of the discriminator D.

5. The transfer learning method based on the correlation between global features and intrinsic features according to claim 4, characterized in that: In the adversarial generation step, Euclidean distance is used to constrain the feature distribution of the connecting samples to lie between the source and target domains. In Formula 1, Represents the Euclidean distance-based constraints between the original samples and the connecting samples in the source domain; In Formula 2, Represents the Euclidean distance-based constraint between the original samples and the connecting samples in the target domain.

6. The transfer learning method based on the correlation between global features and intrinsic features according to claim 3, characterized in that: Perform the adversarial training steps according to Formulas 3 and 4; --Formula 3; --Formula 4; in: y represents the real label; x represents the original sample; Indicates connecting samples; C represents the classifier; This represents the original sample from the source domain with labels. This represents the original sample of the target domain with labels; This represents a sample connecting the source and target domains. Represents cross-entropy; These represent the model parameters of classifier C; Let C represent the loss function of the classifier C; Represents the expected value of the function; This indicates adversarial training; The classification loss function is related to the connecting samples.

7. The transfer learning method based on the correlation between global features and intrinsic features according to claim 6, characterized in that: It also includes using classifier C and discriminator D to obtain labeled raw samples. and The loss; The loss of classifier C is calculated from the original sample x and the adaptive sample x. composition; This includes using discriminator D to obtain unlabeled raw samples. The loss is calculated using formulas five and six. --Formula 5; --Formula Six; in: The loss function of the discriminator D is represented; These represent the model parameters of the discriminator D; D represents the discriminator D; This represents all original samples containing both the source and target domains; This represents the concatenated sample corresponding to all original samples containing both the source and target domains. Represents the expectation of the function; This represents a single original sample belonging to the source domain; This represents a single original sample belonging to the target domain; This represents a single connected sample belonging to the source domain; This represents a single connection sample belonging to the target domain; This indicates adversarial training; This represents the discriminant loss function associated with the connecting samples.

8. The transfer learning method based on the correlation between global features and intrinsic features according to claim 2, characterized in that: In step E, the extraction component includes processing the labeled source domain based on the depth features of the sample. and unselected samples Perform unsupervised clustering; After clustering, samples within the same component contain at least one label.

Images