Incremental Passive Domain Adaptive Method and Device Based on Residual Feature Modulation

Abstract

This invention provides a class-increment passive domain adaptive method and apparatus based on residual feature modulation. The method includes: acquiring target domain data to be identified; inputting the target domain data into a pre-trained classification model, using the pre-trained classification model to calculate the class of the target domain data and outputting corresponding pseudo-labels; the pre-trained classification model includes: a pre-trained source domain feature extractor and a residual target modulation module; the pre-trained source domain feature extractor is trained using source domain data; the residual target modulation module is obtained by adding an attention mechanism and residual connections to the pre-trained source domain feature extractor; updating the pseudo-labels of the classes using the nearest neighbor method to obtain updated class labels; and updating the network parameters of the residual target modulation module using the updated class labels to complete the class increment feature modulation. This improves the learning effect of the pre-trained classification model on the target domain data and reduces the maintenance cost of the pre-trained classification model.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and machine learning technology, specifically relating to a method and apparatus for incremental passive domain adaptation based on residual feature modulation. Background Technology

[0002] In the fields of artificial intelligence and machine learning, Source-Free Domain Adaptation (SFDA) is an important research direction. It focuses on how to transfer a model trained on a source domain to a target domain without labeled information, while maintaining or improving the model's performance in the target domain. The challenge of SFDA lies in the lack of access to source domain data, or even the unknown distribution of the source domain. In this situation, the model must learn solely from unlabeled data in the target domain. This necessitates a stronger generalization ability and better adaptability in the model, enabling it to effectively adapt to the target domain even without source domain information.

[0003] Traditional unsupervised domain adaptation (UDA) is a key technique for achieving cross-domain knowledge transfer. The main principle of UDA is to leverage existing knowledge from the source domain dataset, typically assuming access to labeled data in the source domain and unlabeled data in the target domain. By utilizing this data, the model attempts to learn feature representations that can be transferred from the source domain to the target domain, reducing the distributional differences between the two domains and helping the model achieve better performance in the target domain. The UDA process begins with data preparation, i.e., collecting labeled data from the source domain and unlabeled data from the target domain. However, in some cases, obtaining labeled data from the source domain is expensive or difficult. Next, a base model is pre-trained on the source domain data to learn feature representations. This step often requires significant computational resources, which is impractical in resource-constrained environments. Subsequently, adversarial training or domain-invariant learning methods are used to minimize the distributional differences between the source and target domains. If the distributional differences between the source and target domains are very large, it is difficult to completely align the feature distributions using UDA alone, thus affecting the model's performance in the target domain. Finally, transfer learning is performed to adapt the model to the target domain. Finally, unsupervised evaluation metrics are used to measure the model's performance in the target domain.

[0004] However, in dynamic and continuous real-world environments, data is constantly generated in a streaming manner, and new categories may be continuously created and included. UDA methods are prone to catastrophic forgetting when faced with new category problems. This limits the effectiveness of the model in practical applications, resulting in weak model generalization ability and increasing the complexity of model maintenance and updates. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, this invention provides a method and apparatus for incremental passive domain adaptation based on residual feature modulation.

[0006] The technical problem to be solved by this invention is achieved through the following technical solution:

[0007] In a first aspect, the present invention provides a method for incremental passive domain adaptation based on residual feature modulation, comprising:

[0008] Acquire the target domain data to be identified;

[0009] The target domain data to be identified is input into a pre-trained classification model. The pre-trained classification model is used to calculate the category of the target domain data to be identified and output the corresponding pseudo-label. The pre-trained classification model includes a pre-trained source domain feature extractor and a residual target modulation module. The pre-trained source domain feature extractor is trained using source domain data. The residual target modulation module is obtained by adding an attention mechanism and residual connections to the pre-trained source domain feature extractor.

[0010] The category pseudo-labels are updated using the nearest neighbor method to obtain the updated category labels;

[0011] The network parameters of the residual target modulation module are updated by updating the category labels, thus completing the class increment of feature modulation.

[0012] Optionally, the target domain data to be identified is input into a pre-trained classification model. The pre-trained classification model is then used to calculate the category of the target domain data and output the corresponding pseudo-label, including:

[0013] The first feature is obtained by using a pre-trained source domain feature extractor to extract the first feature from the target domain data to be identified.

[0014] The residual target modulation module is used to extract the second feature from the target domain data to be identified, thus obtaining the second feature.

[0015] The third feature is obtained by residual connection of the first and second features;

[0016] The third feature is used to calculate the category, and the corresponding pseudo-label is obtained.

[0017] Optionally, the pre-trained source domain feature extractor is used to freeze the network structure for parameters during the class increment process for target domain data.

[0018] Optionally, the category pseudo-labels are updated using the nearest neighbor method, resulting in updated category labels including:

[0019] Multiple class prototype features corresponding to the third feature are obtained through the nearest neighbor method;

[0020] Multiple class labels are obtained by performing category calculations on multiple class prototype features;

[0021] The class label with the highest probability among multiple class labels is used as the updated class label.

[0022] Optionally, the network parameters of the residual target modulation module are updated using the updated category labels to complete the class increment of feature modulation, including:

[0023] By using updated category labels and a preset loss function, the parameters of the residual target modulation module are updated under the action of the backpropagation algorithm.

[0024] When the preset loss function reaches the preset convergence condition, the feature modulation class increment is completed.

[0025] Optionally, the default loss function is the sum of cross-entropy loss and contrastive representation enhancement loss.

[0026] Optionally, the residual target modulation module is equipped with a modulation module;

[0027] The modulation module is configured with the following layers in sequence: a 1x1 first convolutional layer, a 3x3 second convolutional layer, a 1x1 third convolutional layer, and a 1x1 fourth convolutional layer.

[0028] The first convolutional layer and the second convolutional layer are connected in series to form the first connection;

[0029] The first convolutional layer, the third convolutional layer, and the fourth convolutional layer are connected in parallel.

[0030] Optionally, the pre-trained classification model may also include: a contrastive representation enhancement module;

[0031] The contrast representation enhancement module is also used to perform data augmentation processing on the target domain data to be identified, so as to obtain enhanced target domain data;

[0032] We use augmented target domain data to perform class incremental reinforcement training on the pre-trained classification model.

[0033] Secondly, the present invention provides a quasi-incremental passive domain adaptive device based on residual feature modulation, comprising: a processor, a storage medium, and a bus. The target domain data storage medium stores machine-readable instructions executable by the target domain data processor. When the quasi-incremental passive domain adaptive device based on residual feature modulation is running, the target domain data processor and the target domain data storage medium communicate via the bus. The target domain data processor executes the machine-readable instructions to perform the steps of the target domain data method described in the first aspect above.

[0034] Thirdly, the present invention provides a storage medium on which a computer program is stored. When the computer program for identifying the target domain data is run by a processor, it performs the steps of the method for identifying the target domain data as described in the first aspect above.

[0035] This invention provides a class-increment passive domain adaptive method and apparatus based on residual feature modulation. The class-increment passive domain adaptive method based on residual feature modulation includes: acquiring target domain data to be identified; inputting the target domain data to be identified into a pre-trained classification model, using the pre-trained classification model to calculate the category of the target domain data and outputting the corresponding pseudo-label; the pre-trained classification model includes: a pre-trained source domain feature extractor and a residual target modulation module; the pre-trained source domain feature extractor is trained using source domain data; the residual target modulation module is obtained by adding an attention mechanism and residual connections to the pre-trained source domain feature extractor; updating the pseudo-labels of the categories using the nearest neighbor method to obtain updated category labels; and updating the network parameters of the residual target modulation module using the updated category labels to complete the class increment of feature modulation. In this invention, by setting up a pre-trained source domain feature extractor and a residual target modulation module, and using both modules together for category calculation, not only can the generalization ability of the pre-trained classification model to source domain data be preserved, but effective discrimination of target domain data can also be achieved. Secondly, by gradually introducing target domain data, the pre-trained classification model can maintain its stability when facing continuously changing data streams. Finally, by continuously updating the pseudo-labels of the categories using the nearest neighbor method, the stability and reliability of the pseudo-labels can be improved, thereby enhancing the learning effect of the pre-trained classification model on the target domain data. Since the entire process involves only a few changes to the model structure, the cost of parameter training is reduced, thus lowering the maintenance cost of the pre-trained classification model.

[0036] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating a quasi-incremental passive domain adaptive method based on residual feature modulation according to an embodiment of the present invention.

[0038] Figure 2 A schematic diagram of the overall process of an incremental passive domain adaptive method based on residual feature modulation provided in another embodiment of the present invention;

[0039] Figure 3 This is a schematic diagram of a quasi-incremental passive domain adaptive device based on residual feature modulation, provided in an embodiment of the present invention. Detailed Implementation

[0040] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0041] To improve the generalization ability of the model and enhance its learning performance on target domain data, this invention provides a method for incremental passive domain adaptation based on residual feature modulation. Figure 1 This is a flowchart illustrating a quasi-incremental passive domain adaptive method based on residual feature modulation, provided in an embodiment of the present invention. Figure 1 As shown, it includes:

[0042] S101. Obtain the target domain data to be identified.

[0043] It should be noted that in this embodiment, the target domain data and the source domain data are data from similar domains. For example, if the target domain data is license plate data for rainy days, then the source domain data is license plate data for sunny days. That is, the output categories of the target domain data and the source domain data remain the same.

[0044] Furthermore, since the pre-trained classification model is continuously self-updating, the target domain data to be identified can be considered both training data and data to be identified. By gradually introducing target domain data, the pre-trained classification model can maintain its stability when faced with continuously changing data streams.

[0045] S102. Input the target domain data to be identified into the pre-trained classification model, use the pre-trained classification model to calculate the category of the target domain data to be identified and output the corresponding category pseudo-label.

[0046] Optionally, step S102 may specifically include:

[0047] The first feature is obtained by using a pre-trained source domain feature extractor to extract the first feature from the target domain data to be identified.

[0048] The residual target modulation module is used to extract the second feature from the target domain data to be identified, thus obtaining the second feature.

[0049] The third feature is obtained by residual connection of the first and second features;

[0050] The third feature is used to calculate the category, and the corresponding pseudo-label is obtained.

[0051] The pre-trained classification model includes a pre-trained source domain feature extractor and a residual target modulation module. The pre-trained source domain feature extractor is trained using source domain data. The residual target modulation module is obtained by adding an attention mechanism and residual connections to the pre-trained source domain feature extractor.

[0052] In this embodiment of the application, the pre-trained classification model further includes a classifier, which is used to calculate the category of the generated third feature and finally obtain the corresponding category pseudo-label.

[0053] Furthermore, in this embodiment, for each category in the target dataset, the cumulative sum of predicted probabilities of all predicted results (pseudo-labels) can be calculated and normalized. This step converts the predicted probabilities into a relative metric, facilitating comparison of the importance of different categories, and simultaneously determines which categories are the target categories by setting a threshold (α). If the normalized cumulative predicted value of a category exceeds the threshold, then that category is considered the target category for the current learning step. Once the target category is determined, it can be used as the pseudo-label for the current target domain data to be identified.

[0054] Optionally, the residual target modulation module is provided with a modulation module; the modulation module is provided with the following in sequence: a 1x1 first convolutional layer, a 3x3 second convolutional layer, a 1x1 third convolutional layer and a 1x1 fourth convolutional layer;

[0055] The first convolutional layer and the second convolutional layer are connected in series to form the first connection;

[0056] The first convolutional layer, the third convolutional layer, and the fourth convolutional layer are connected in parallel.

[0057] It should be noted that in this embodiment, the 1x1 first convolutional layer is mainly used to reduce the number of channels in the feature map while maintaining the spatial dimension, which helps to reduce computational complexity. The 3x3 second convolutional layer is used to capture local spatial information and enhance the expressive power of the features.

[0058] Specifically, the features obtained by the pre-trained source domain feature extractor are filtered sequentially through the first and second convolutional layers to obtain a latent feature map (z). The latent feature map is then mapped through the third and fourth convolutional layers to generate the third and fourth feature maps (zq and zk). Softmax modulation is performed on zq and zk to obtain the modulated feature map (zm). The original feature map is added to the modulated feature map (by performing a residual connection, i.e., adding at the feature element level) to obtain the target feature map (zt), which is then used as the third feature.

[0059] It should be noted that the residual connection structure helps prevent the vanishing gradient problem and allows the model to learn new feature representations more effectively during training. Furthermore, since the residual target modulation module has a relatively small number of parameters, it also helps reduce the training complexity and improve training efficiency.

[0060] Optionally, the pre-trained source domain feature extractor is used to freeze the network structure for parameters during the class increment process for target domain data.

[0061] In this embodiment, the pre-trained source domain feature extractor is configured as a parameter-frozen network structure during the class increment process for target domain data. That is, the weights of the pre-trained source domain feature extractor remain unchanged during the continuous optimization of the pre-trained classification model, thus avoiding the loss of source domain knowledge. Simultaneously, the parameters of the residual target modulation module are trainable and can be fine-tuned based on the target domain data to achieve adaptation to the target domain.

[0062] S103. Update the category pseudo-labels using the nearest neighbor method to obtain the updated category labels.

[0063] Optionally, step S103 may specifically include:

[0064] Multiple class prototype features corresponding to the third feature are obtained through the nearest neighbor method;

[0065] Multiple class labels are obtained by performing category calculations on multiple class prototype features;

[0066] The class label with the highest probability among multiple class labels is used as the updated class label.

[0067] It should be noted that the pseudo-labels of the categories are updated using the nearest neighbor method. Specifically, for each feature corresponding to the target domain data i to be identified, its k most similar neighbors in the feature space are found, and the pseudo-label of i is updated based on the pseudo-labels of these neighbors. That is, if most of the most similar neighbors of the feature of i belong to the same category (the class label with the highest probability distribution), then the class label with the highest probability distribution is used as the new pseudo-label of i. This process helps to correct errors in the category pseudo-labels, and the use of class prototypes makes the correction process more robust.

[0068] Furthermore, the multiple class prototype features can specifically be feature data obtained from the input data of the target domain data to be identified through the nearest neighbor method. Moreover, the pre-trained classification model in this application is continuously self-updating, constantly optimizing its network parameters based on the acquired target domain data.

[0069] Furthermore, during the iteration process, the performance of the pre-trained classification model on the target dataset can be continuously monitored. For example, metrics such as accuracy and loss function can be used to evaluate the effect of pseudo-label correction.

[0070] If performance metrics show no significant improvement in the pre-trained classification model's performance, or if the distribution of pseudo-labels remains unstable, the iteration strategy can be adjusted appropriately. For example, changing the method for calculating class prototypes, adjusting the number of nearest neighbors, or modifying the initial pseudo-label generation threshold (α).

[0071] By employing the aforementioned class pseudo-label correction strategy, this invention effectively improves the quality of class pseudo-labels in the target domain data, reducing model training problems caused by noise and mislabels. This not only helps the model better adapt to the target domain but also reduces catastrophic forgetting and improves the pre-trained classification model's ability to recognize new categories. Furthermore, this strategy can adapt to new categories in the target data stream, achieving the goal of incremental class learning.

[0072] S104. Update the network parameters of the residual target modulation module using the updated category label to complete the class increment of feature modulation.

[0073] This invention provides a class-incremental passive domain adaptive method based on residual feature modulation, comprising: acquiring target domain data to be identified; inputting the target domain data to be identified into a pre-trained classification model, using the pre-trained classification model to calculate the category of the target domain data to be identified and outputting the corresponding category pseudo-label; the pre-trained classification model includes: a pre-trained source domain feature extractor and a residual target modulation module; the pre-trained source domain feature extractor is trained using source domain data; the residual target modulation module is obtained by adding an attention mechanism and residual connections to the pre-trained source domain feature extractor; updating the category pseudo-labels using the nearest neighbor method to obtain updated category labels; and updating the network parameters of the residual target modulation module using the updated category labels to complete the class increment of feature modulation. In this embodiment of the invention, by setting up a pre-trained source domain feature extractor and a residual target modulation module, and using both modules for category calculation, not only can the generalization ability of the pre-trained classification model to source domain data be preserved, but effective discrimination of target domain data can also be achieved. Secondly, by gradually introducing target domain data, the pre-trained classification model can maintain its stability when facing continuously changing data streams. Finally, by continuously updating the pseudo-labels of the categories using the nearest neighbor method, the stability and reliability of the pseudo-labels can be improved, thereby enhancing the learning effect of the pre-trained classification model on the target domain data. Since the entire process involves only a few changes to the model structure, the cost of parameter training is reduced, thus lowering the maintenance cost of the pre-trained classification model.

[0074] Step S104 may specifically include:

[0075] By using updated category labels and a preset loss function, the parameters of the residual target modulation module are updated under the action of the backpropagation algorithm.

[0076] When the preset loss function reaches the preset convergence condition, the feature modulation class increment is completed.

[0077] Optionally, the default loss function is the sum of cross-entropy loss and contrastive representation enhancement loss.

[0078] It should be noted that contrastive representation enhancement loss is particularly suitable for unsupervised domain adaptation scenarios because it does not rely on labeled data but guides model learning through the similarity and differences between data. Therefore, contrastive representation enhancement loss can effectively improve the quality of feature representations in the target domain, enabling the model to have better generalization and discriminative ability in the target domain. Furthermore, contrastive representation enhancement loss can also adapt to new categories in the target domain data stream, achieving the goal of incremental class learning. Moreover, the model training method combining cross-entropy loss and contrastive representation enhancement loss in this embodiment of the invention further enables the model to maintain its memory of old categories while adapting to new categories.

[0079] Optionally, the pre-trained classification model may also include: a contrastive representation enhancement module;

[0080] The contrast representation enhancement module is also used to perform data augmentation processing on the target domain data to be identified, so as to obtain enhanced target domain data;

[0081] We use augmented target domain data to perform class incremental reinforcement training on the pre-trained classification model.

[0082] It should be noted that, in this embodiment, the contrast representation enhancement module can transform the target domain data to be identified using various data augmentation techniques, such as random rotation, scaling, cropping, and color transformation. These enhancement techniques generate variants of the target domain data, increasing data diversity and helping the pre-trained classification model learn more robust feature representations.

[0083] To provide a comprehensive overview of the incremental passive domain adaptive method based on residual feature modulation of the present invention, this embodiment provides a general description of the steps of the method. Specifically, Figure 2 This is a schematic diagram of the overall process of an incremental passive domain adaptive method based on residual feature modulation, which is provided in another embodiment of the present invention.

[0084] like Figure 2As shown, the process begins by acquiring a pre-trained source model. Based on this model, the target data stream is received, and the pre-trained source model is used for label discovery (in cases where the output categories of the source and target data are inconsistent) and pseudo-label prediction. The predicted pseudo-labels are used to calculate cumulative prediction probabilities, and pseudo-labels with prediction probabilities greater than a preset threshold α are taken as the target category (target pseudo-label). Then, feature representation modulation (RTM) is performed. Specifically, the modulation module (RTM) is added to the pre-trained source model, freezing the pre-trained source model parameters while fine-tuning the RTM parameters. Next, iterative correction of the pseudo-labels is performed (including calculating class prototypes, using the nearest neighbor method to correct the target pseudo-labels based on the class prototypes, and using the correction result as the corrected pseudo-label). Simultaneously, contrastive representation enhancement (adding a projection head) is used for data augmentation. After data augmentation, contrastive learning loss is optimized. Throughout the model training process, the sum of cross-entropy loss and contrastive learning loss is used as the overall loss function of the classification model, allowing for continuous self-optimization during use. For new categories, the above steps can be executed multiple times to further perform class incremental learning of the classification model.

[0085] The method provided in this invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc. No limitation is made herein; any electronic device that can implement this invention falls within the protection scope of this invention.

[0086] Based on the same inventive concept, embodiments of the present invention also provide a quasi-incremental passive domain adaptive device based on residual feature modulation. Figure 3 This is a schematic diagram of a quasi-incremental passive domain adaptive device based on residual feature modulation, provided in an embodiment of the present invention. It includes a processor 210, a storage medium 220, and a bus 230. The storage medium 220 stores machine-readable instructions executable by the processor 210. When the quasi-incremental passive domain adaptive device based on residual feature modulation is running, the processor 210 communicates with the storage medium 220 via the bus 230. The processor 210 executes the machine-readable instructions to perform the steps of the above-described method embodiment. Specific implementation methods and technical effects are similar and will not be repeated here.

[0087] The storage medium may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the storage medium may also be at least one storage device located remotely from the aforementioned processor.

[0088] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0089] The present invention also provides a computer-readable storage medium. A computer program is stored in the computer-readable storage medium, and when executed by a processor, the computer program implements the steps of any of the above-described incremental passive domain adaptive methods based on residual feature modulation.

[0090] It should be noted that the terms "first," "second," etc., are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The implementations described in the following exemplary embodiments do not represent all implementations consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure.

[0091] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0092] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings and the disclosure in carrying out the claimed invention. In the description of the invention, the word "comprising" does not exclude other components or steps, "a" or "an" does not exclude a plurality, and "a plurality" means two or more, unless otherwise explicitly specified. Furthermore, while different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce good results.

[0093] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for adaptive learning in a passive domain similar to incremental modulation based on residual feature modulation, characterized in that, include: Acquire the target domain data to be identified; The target domain data to be identified is input into a pre-trained classification model, and the pre-trained classification model is used to calculate the category of the target domain data to be identified and output the corresponding category pseudo-label. The pre-trained classification model includes: a pre-trained source domain feature extractor and a residual target modulation module; the pre-trained source domain feature extractor is trained using source domain data; the residual target modulation module is obtained by adding an attention mechanism and residual connections to the pre-trained source domain feature extractor. The category pseudo-labels are updated using the nearest neighbor method to obtain the updated category labels; The network parameters of the residual target modulation module are updated using the updated category label to complete the class increment of feature modulation; The target domain data to be identified is license plate data from rainy days, and the source domain data is license plate data from sunny days; the target domain data to be identified and the source domain data have the same output category in the pre-trained classification model. The residual target modulation module is equipped with a modulation module; The modulation module is provided with the following layers in sequence: a 1x1 first convolutional layer, a 3x3 second convolutional layer, a 1x1 third convolutional layer, and a 1x1 fourth convolutional layer. The first convolutional layer and the second convolutional layer are connected in series to form a first connection; The first connection, the third convolutional layer, and the fourth convolutional layer are connected in parallel; The pre-trained classification model further includes: a contrastive representation enhancement module; The contrast representation enhancement module is also used to perform data enhancement processing on the target domain data to be identified, to obtain enhanced target domain data; The pre-trained classification model is subjected to class incremental reinforcement training using the enhanced target domain data.

2. The incremental passive domain adaptive method based on residual feature modulation according to claim 1, characterized in that, The step of inputting the target domain data to be identified into a pre-trained classification model, using the pre-trained classification model to calculate the category of the target domain data to be identified, and outputting the corresponding category pseudo-label includes: The pre-trained source domain feature extractor is used to extract the first feature from the target domain data to be identified, thereby obtaining the first feature. The residual target modulation module is used to extract the second feature from the target domain data to be identified, thereby obtaining the second feature. The first feature and the second feature are residually joined to obtain the third feature; The category is calculated using the third feature to obtain the corresponding category pseudo-label.

3. The incremental passive domain adaptive method based on residual feature modulation according to claim 1, characterized in that, The pre-trained source domain feature extractor is a network structure with frozen parameters during the class increment process for target domain data.

4. The incremental passive domain adaptive method based on residual feature modulation according to claim 2, characterized in that, The step of updating the category pseudo-label using the nearest neighbor method to obtain the updated category label includes: The nearest neighbor method is used to obtain multiple class prototype features corresponding to the third feature; Multiple class labels are obtained by performing category calculations on the multiple class prototype features; The class label with the highest probability distribution among the multiple class labels is used as the updated class label.

5. The incremental passive domain adaptive method based on residual feature modulation according to claim 1, characterized in that, The step of updating the network parameters of the residual target modulation module using the updated category label to complete the class increment of feature modulation includes: Using the updated category label and the preset loss function, the parameters of the residual target modulation module are updated under the action of the backpropagation algorithm; When the preset loss function reaches the preset convergence condition, the feature modulation class increment is completed.

6. The incremental passive domain adaptive method based on residual feature modulation according to claim 5, characterized in that, The preset loss function is the sum of cross-entropy loss and contrast representation enhancement loss.

7. A quasi-incremental passive domain adaptive device based on residual feature modulation, characterized in that, include: The device includes a processor, a storage medium, and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the residual feature modulation-based incremental passive domain adaptive device is running, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the steps of the method as described in any one of claims 1-6.

8. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, performs the steps of the method as described in any one of claims 1-6.

Images