Class-domain incremental learning method for power distribution network environment ground feature element extraction model

Abstract

The invention provides a class-domain incremental learning method for a ground feature element extraction model of a power distribution network environment, and the method achieves cross-domain knowledge accumulation through feature synthesis: firstly, introducing a temporary style encoder in a training process, and taking a backbone network of the ground feature element extraction model as a content encoder; effective decoupling of disaster remote sensing image contents and styles is realized by applying orthogonal constraint to the outputs of the two, so that specific style features of a domain are maintained in the whole incremental learning stage; projecting the stored old domain style to the new domain content to generate old domain features for supporting old knowledge maintenance, realizing extension of a new category to an old domain, and promoting old category learning on a new domain image; in addition, style randomization strategies of category perception are integrated, so that the image feature decoupling capability is enhanced, and the quality of synthetic features is improved. The generalization and application capability of the ground feature element extraction model in complex and diversified power distribution network scenes is enhanced, and the method is suitable for power distribution network disaster monitoring tasks.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology and relates to a semantic segmentation method for remote sensing images, specifically a class-domain incremental learning method for extracting ground features in a power distribution network environment. Background Technology

[0002] The development of deep learning has driven improvements in the accuracy of remote sensing cognition of disaster elements. However, this high-precision performance is usually based on the assumption of a closed world, that is, the model will only encounter predefined categories of ground features during the testing phase, and the test images and training images are independent and identically distributed. Real-world applications struggle to meet this closed-world assumption for two main reasons: Firstly, power distribution network equipment is numerous, covers a wide area, and is dispersed; similar targets in different power distribution network scenarios may exhibit significant differences, and even new target categories may emerge. Secondly, the iteration of remote sensing Earth observation technologies gives rise to new data sources, and cognitive requirements change accordingly. Therefore, ground feature extraction models built based on static training paradigms can only meet specific cognitive requirements in specific scenarios. Once the application scenario or cognitive requirements change, the model's performance is difficult to guarantee, and it may even become unusable. This severely hinders the wide-area and long-term application of ground feature extraction models. To adapt to open environments and meet dynamic needs, ground feature extraction models, based on the infinite flow of disaster remote sensing data, should possess the ability to continuously absorb new knowledge while retaining and even optimizing old knowledge, thus constantly evolving. The simplest model evolution strategy is to fine-tune the extraction model using both new and old data, achieving high performance across multiple training datasets for feature extraction tasks through joint optimization. However, due to factors such as privacy protection and storage limitations, training data for older tasks (or requirements) may be unavailable. Training the model solely with data from new tasks can easily lead to catastrophic forgetting of knowledge from older tasks, resulting in performance degradation on those tasks or datasets. Therefore, how to guide the extraction model to continuously learn new knowledge while effectively maintaining existing knowledge when older data is unavailable is a crucial technical problem that needs to be solved for the commercial application of feature extraction models. Summary of the Invention

[0003] To address the problem that remote sensing cognitive models of typical disaster elements in power distribution networks lack continuous learning capabilities and are difficult to support long-term and large-scale applications, this invention proposes a class-domain incremental learning method for extracting ground features in power distribution networks. This method achieves cross-domain knowledge accumulation through feature synthesis, enhancing the generalization and application capabilities of the extraction model in complex and diverse power distribution network scenarios, and is suitable for long-term and wide-area power distribution network disaster monitoring tasks.

[0004] The class-domain incremental learning method for extracting ground features in power distribution network environments provided by this invention has the following main improvements: 1. A temporary style encoder is added to the ground feature extraction model of the encoder-decoder architecture, and the backbone network of the extraction model is regarded as a content encoder. The style encoder is used to extract the style features of the image, while the content encoder is used to extract the content features of the image. By applying orthogonal constraints to the two types of feature representations, the feature decoupling of content and style of disaster remote sensing images is achieved.

[0005] 2. Based on the feature decoupling framework, a style memory is maintained throughout the incremental learning process to store style features from different domains. The style features of a specific domain are the average of the style information of all training images in the current domain. In subsequent learning stages, the stored style features of the specific domain are fused with the content features of the new domain images to reconstruct the image representation of the old domain, supporting the playback of knowledge from the old domain.

[0006] 3. Four specially designed loss functions are used to guide model training: learning new categories in new / old domains, maintaining old category knowledge learned in the old domain, and extending old category knowledge to the new domain, thereby achieving cross-domain category knowledge accumulation.

[0007] 4. The performance of the proposed method depends on the effective decoupling of content features and style features, as the decoupling effect directly affects the synthesis quality of old domain features. However, simply applying orthogonal constraints is insufficient to guarantee the effective separation of content features and style features; that is, content features should reflect the high-level semantic information of the image, while style features should reflect the low-level appearance information. To address this issue, the proposed method introduces a category-aware style randomization strategy. By generating randomized stylized versions of the training images and maintaining consistency between the stylized images and the original images' content features, the method guides the content encoder to focus on semantic information extraction rather than style changes.

[0008] The principle of this invention is to utilize content-style feature decoupling and style randomization guidance, combined with the old domain synthetic feature replay mechanism, to achieve cross-domain category incremental learning and knowledge preservation, thereby endowing the ground feature extraction model with continuous learning capabilities.

[0009] The specific implementation of the method of the present invention includes the following steps: Content features and style features of images are extracted by a ground feature extraction model encoder and a temporary style encoder, respectively. Orthogonal constraints are applied to the two types of feature representations in each incremental learning stage to decouple content and style. The style features of all training images in the current domain are aggregated to obtain domain-specific average style features and stored in the style memory. The instance style features or domain-specific average style features are projected onto the content features to synthesize a complete image representation, which is then input into the decoder to generate the classification prediction results of the learning stage. The average style features of each old domain in the style memory are projected onto the content features of the new domain image to synthesize the feature representation of the old domain image, and then input into the decoder to generate pseudo-labels of the old categories for the new domain image; based on the ground truth labels of the new categories and the pseudo-labels of the old categories, a loss function is designed to achieve cross-domain category knowledge accumulation. For new domain images, the same style randomization operation is applied to targets belonging to the same category, and the processing results of each category are combined to obtain a stylized image. By constraining the consistency loss between the stylized image and the original image in terms of content features, the quality of the synthesized features is improved.

[0010] Furthermore, in the Each incremental learning phase will use the remote sensing images of the current domain. Input content encoders respectively and style encoder This generates corresponding depth features. , ,in Indicates spatial dimension, This refers to the channel dimension; the orthogonal constraint is achieved by minimizing the cosine similarity between content features and style features, specifically expressed as:

[0011] in, This represents the training dataset for the current learning phase. It is cosine similarity. It is global average pooling.

[0012] Furthermore, by driving volume The exponential moving average method is used to obtain style features of a specific domain. It is iteratively updated during training, and its update formula is as follows:

[0013] in Indicates the first In each training iteration, from the input The extracted style features, and the domain-specific style features in the incremental learning phase, are denoted as... , The total number of iterations during the learning phase will be used to store this style feature in the style memory.

[0014] Furthermore, a random fusion strategy is adopted, using random sampling probability. The choice between using domain-specific style features or style features extracted from the current batch for fusion is thus made during the training process. The calculation is as follows:

[0015] in, For random selection or Indicates the first In each training iteration, from the input Extracted content features Indicates the first In each training iteration, from the input Extracted style features Indicates domain-specific style characteristics, This indicates the fusion module, which consists of a convolutional layer, an AdaIN layer, and a ReLU activation layer.

[0016] Furthermore, adopting the first The extraction model generated during each learning stage produces pseudo-labels for the old categories:

[0017] in This indicates that the extraction model in the previous learning stage is in the [number]th [stage]. The predicted classification probability of a new domain image under a domain-specific style. Represents domain-specific style characteristics; , , These are the content encoder, fusion module, and decoder from the previous learning stage.

[0018] Furthermore, based on the new category truth labeling and old category pseudo-labels Four loss functions were designed, the first of which is the new domain-new class learning loss function. The second old-new-class expansion loss function is used to supervise learning of new categories in a new domain. The third new-domain-old-class adaptation loss function is used to guide the model to extend knowledge of new categories to the old domain. The fourth old-domain-old-class preservation loss function is used for old-class adaptation on the new domain. Used to ensure that the current model does not forget the old category knowledge learned in the old domain.

[0019] Furthermore, the first new domain-new category learning loss function The function form is as follows:

[0020] in,

[0021] It is a decoder In receiving new domain features The resulting classification probability; Represents the model at pixel position Category The predicted probability, For remote sensing images, This represents the training dataset for the current learning phase; It is the expanded category space, containing all the old categories. and background categories ; Second old domain - new category extended loss function The function form is:

[0022] in, It is by Based on intermediate variables The calculated value is used to solve the background offset problem. It is the decoder In receiving synthesized old domain features The resulting predictions; The third new domain-old category adaptation loss function The calculation process is as follows: First, the old model predictions are filtered using the following strategy to generate reliable pseudo-labels:

[0023] Combining truth labels of new categories Optimize:

[0024] Filtering function right Perform the screening. Finally... Calculated by the following formula:

[0025] Similarly, These are the modified classification probabilities to address the background offset issue:

[0026] .

[0027] The fourth old domain-old category preservation loss function The calculation form is as follows:

[0028] It is by The calculation results are as follows:

[0029] Based on the four loss functions mentioned above, the cross-domain category knowledge accumulation function is:

[0030] = = This is an adjustable hyperparameter used to balance the weights of various loss terms.

[0031] Furthermore, in the style randomization process, convolutional layers with random sampling weights are first used to enrich texture features, and then randomized style statistics are used to replace the original statistics, thereby achieving style diversification.

[0032] Furthermore, the calculation of the content consistency constraint loss includes: inputting the content features of the stylized image, the content features of the original image, and the style features of the original image into the reconstruction module to generate a reconstructed image, and achieving content consistency constraint by minimizing the difference between the reconstructed image and the original image.

[0033] The present invention also provides a class-domain incremental learning system for a distribution network environment feature extraction model, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute the class-domain incremental learning method for a distribution network environment feature extraction model as described in the above technical solution.

[0034] This invention achieves cross-domain knowledge accumulation through feature synthesis: First, a temporary style encoder is introduced during training, and the backbone network of the feature extraction model is used as the content encoder. By applying orthogonal constraints to the outputs of both, effective decoupling of content and style in disaster remote sensing images is achieved, thus maintaining domain-specific style features throughout the incremental learning phase. Then, stored old domain styles are projected onto new domain content to generate old domain features, supporting the preservation of old knowledge, enabling the expansion of new categories to the old domain, and promoting the learning of old categories on new domain images. Furthermore, a category-aware style randomization strategy is integrated to enhance image feature decoupling capabilities and improve the quality of synthesized features. This invention enhances the generalization and application capabilities of feature extraction models in complex and diverse power distribution network scenarios, and is suitable for long-term, wide-area power distribution network disaster monitoring tasks. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the overall framework of an embodiment of the present invention.

[0036] Figure 2 This is a schematic diagram illustrating the calculation of four customized losses for cross-domain category knowledge accumulation during incremental learning. Detailed Implementation

[0037] This invention provides a joint incremental learning method for categories and domains in a model for extracting ground features in a power distribution network environment. To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings.

[0038] The class-domain incremental learning method for extracting ground features in power distribution network environments provided in this embodiment of the invention includes the following steps: Image feature decoupling and synthesis: First, content features and style features of the image are extracted using an existing land feature extraction model encoder and a temporary style encoder, respectively. Orthogonal constraints are applied to the two types of feature representations to achieve decoupling between content and style. During training, an exponentially weighted average algorithm is used to aggregate the style features of all training images in the current domain to obtain domain-specific average style features, which are stored in a style memory. Subsequently, a lightweight fusion module projects the instance style features or domain-specific average style features onto the content features to synthesize a complete image representation, which is finally input into the decoder to generate classification prediction results. Joint incremental learning of categories and domains: The average style features of each old domain in the style memory are projected onto the content features of the new domain image to synthesize the feature representation of the old domain image. This representation is then input into the decoder of the old model to generate pseudo-labels for the old categories on the new domain image. Based on the ground truth labels of the new categories and the pseudo-labels of the old categories, four loss functions are customized for: learning new categories in the new domain, adapting old categories in the new domain, expanding new categories in the old domain, and maintaining old categories in the old domain, thereby achieving cross-domain category knowledge accumulation. Category-aware style randomization: For images in the new domain, targets belonging to the same category are subjected to the same style randomization operation, and the processing results of each category are combined to obtain a stylized image. By constraining the consistency of content features between the stylized image and the original image, the encoder of the extraction model is prompted to focus more on the semantic information representation of the image, thereby achieving effective decoupling of content and style and further improving the quality of synthesized features. To ensure content consistency while giving discriminative feature learning flexibility, a temporary reconstruction module is introduced. This module requires that the stylized content features, the original content features, and the original style features can reconstruct the original input image, thus establishing a content consistency constraint.

[0039] The implementation of the present invention will be described in detail below: Step 1: In the Each incremental learning phase will use the disaster remote sensing images from the current domain. Input content encoders respectively and style encoder This generates corresponding depth features. , ,in Indicates spatial dimension, It is the channel dimension. An orthogonal constraint is imposed by minimizing the following loss:

[0040] in, This represents the training dataset for the current learning phase. It is cosine similarity. It is global average pooling.

[0041] Step 2: After achieving feature decoupling, by driving the momentum The exponential moving average method is used to obtain style features of a specific domain. It is iteratively updated during training, and its update formula is as follows:

[0042] in Indicates the first In each training iteration, from the input Extracted style features. Domain-specific style features in the incremental learning phase are denoted as... , This represents the total number of iterations for this learning phase. The style feature will be stored in the style memory. .

[0043] Step 3: During the inference phase, since the style encoder is discarded after training, the domain-specific style features are projected onto the content features of the test image to form a complete image representation, which is then input into the decoder to generate the final ground feature extraction result. During the training phase, a random fusion strategy is employed, selecting to fuse either domain-specific style features or style features extracted in the current batch with a 50% probability. This random selection strategy enhances the decoder's generalization ability. Therefore, the input features of the decoder during training... The calculation is as follows:

[0044] in

[0045] This represents the fusion module, which consists of convolutional layers, AdaIN layers, and ReLU activation layers in sequence. This represents the probability of random sampling.

[0046] Step 4: Based on the aforementioned style memory and fusion module By projecting the old domain style features onto the first The content features of the training images during the learning phase can be used to synthesize the old domain image representation, as shown in the following formula:

[0047] in, It is in the The old domain image representation at the learning stage, where the superscript indicates the learning stage. During the training iteration, the first... The content features of the first learning stage were synthesized into the second... Image representation in each learning stage.

[0048] Step 5: Truth Labeling Supervised extraction models can learn new categories in new / old domains, but cannot provide effective supervision for preserving old categories in old domains or adapting old categories in new domains. This is because... In this context, old category targets are labeled as background. To address this issue, the method provided by this invention employs the following approach: The extraction model generated during each learning stage produces pseudo-labels for the old categories:

[0049] in This indicates that the extraction model in the previous learning stage is in the [number]th [stage]. The predicted classification probability of an image in a new domain under a specific style of the current domain. , , These are the content encoder, fusion module, and decoder from the previous learning stage. Because the style memory stores the previous... The style of each learning stage, therefore the old model will generate Each prediction result is denoted as... .

[0050] Step 6: Truth labeling based on the new category and old category pseudo-labels Four loss functions were designed to achieve cross-domain category knowledge accumulation. The first is the new domain-new class learning loss function. Used to supervise the learning of new categories in a new domain, the function has the following form:

[0051] in,

[0052] It is a decoder In receiving new domain features The resulting classification probability. Represents the model at pixel position Category The predicted probability. It is the expanded category space, containing all the old categories. and background categories .

[0053] Step 7: Expand the loss function in the second old domain - new category The function used to guide the model to extend knowledge of new categories to the old domain is as follows:

[0054] in, It is by The formula calculated in step 6 is used to solve the background offset problem. It is the decoder In receiving synthesized old domain features The resulting predictions.

[0055] Step 8: Third New Domain - Old Class Adaptation Loss Function Used for old class adaptation on new domains. Although the old model can provide old class pseudo-labels for the current model's learning. However, these pseudo-labels inevitably contain incorrect predictions. To address this issue, this invention employs the following strategy to filter the predictions of the old model, thereby generating reliable pseudo-labels:

[0056] To further reduce noise in pseudo-labels, this invention incorporates a new category of truth-value pairings. Optimize:

[0057] , and For simplification, the following was omitted. Filtering function right The selection process involves filtering and retaining only predictions from older models with sufficiently high confidence levels. The criteria for this selection include two aspects: peak probability. Above the threshold Or the peak probability in category The middle position is ranked in the top percentile. . This indicates that the pixel is ignored and will not be included in the loss calculation. Ultimately... It can be calculated using the following formula:

[0058] Similarly, These are the modified classification probabilities to address the background offset issue:

[0059] .

[0060] Step 9: Fourth Old Domain - Old Class Preservation Loss Function To ensure the current model does not forget the old category knowledge learned in the old domain, its computational form is as follows:

[0061] It is by The calculation results are as follows:

[0062] Based on the four loss functions mentioned above, the cross-domain category knowledge accumulation function is:

[0063] = = This is an adjustable hyperparameter used to balance the weights of various loss terms.

[0064] Step 10: To improve the quality of synthesized features, this invention proposes category-aware style randomization. Targets belonging to the same category in the input image are subjected to the same style randomization process, and then combined to generate the final stylized output. During style randomization, convolutional layers with randomized sampling weights are first used to enrich texture features, and then randomized style statistics are used to replace the original statistics, thereby achieving style diversification. Targets belonging to the same category in the new domain image are defined as... ,in For category The binary mask. For Style randomization involves two consecutive steps. First, through... Recursive texture modification is performed using random convolutional layers:

[0065] in From the set Mid-sampling is used to achieve sufficiently rich texture variations without compromising semantics. Convolutional weights for each layer... Follows Gaussian distribution kernel size From the set Randomly selected from the options. Then, a style transformation is performed:

[0066] in, and Calculate the standard deviation and mean of the input channel dimension, and the random statistics respectively. and From the standard normal distribution Medium sampling. To reduce potential semantic distortion, this invention performs a weighted combination of the original image and the stylized image for each category, using randomly sampled mixing coefficients. Final stylized image The result obtained by aggregating all categories is:

[0067] Step 11: This invention does not rely on explicit feature alignment, but instead introduces a temporary reconstruction module. This module is jointly trained with the extraction model and is discarded after training. This module uses content features... , and original style characteristics Using the input image, a reconstructed image is generated. A content consistency loss function is then constructed for model optimization:

[0068] in During training, the content consistency loss and orthogonality constraint loss will be jointly optimized to achieve effective decoupling of content features and style features.

[0069] Step 12: Overall Optimization Objective of the Land Feature Extraction Model As shown in the following formula:

[0070] and These are the weighting factors for the content-style orthogonality constraint and the stylization-original content consistency loss, respectively. During the inference phase, the style features of the latest domain are directly projected onto the content features of the test image to construct a complete image representation, which is then input into the decoder for prediction.

[0071] Table 1. Results of the joint incremental experiment on the category and domain datasets of the ISPRS dataset.

[0072] Experimental validation: The method was tested on a joint category and domain incremental experiment constructed using the ISPRS dataset, and the results are shown in Table 1. The experimental results show that, compared with existing methods such as LwS and MiSSNet, the method of this invention exhibits superior performance in all experimental settings, achieving performance improvements of 1.89%, 7.78%, and 4.90%, respectively.

[0073] On the other hand, embodiments of the present invention also provide a class-domain incremental learning system for a distribution network environment feature extraction model, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute the class-domain incremental learning method for a distribution network environment feature extraction model as described in the above technical solution.

[0074] Thirdly, embodiments of the present invention also provide a computer-readable storage medium, including a readable storage medium on which a computer program is stored, wherein when the computer program is executed, it implements the class-domain incremental learning method for extracting ground features in the power distribution network environment as described in the above technical solution.

[0075] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A class-domain incremental learning method for extracting environmental features in power distribution networks, characterized in that, Includes the following steps: Content features and style features of images are extracted by the ground feature extraction model encoder and the temporary style encoder, respectively. Orthogonal constraints are applied to the two types of feature representations in each incremental learning stage to achieve decoupling of content and style. The style features of all training images in the current domain are aggregated to obtain domain-specific average style features and stored in the style memory bank. Instance style features or domain-specific average style features are projected onto content features to synthesize a complete image representation, which is then input into the decoder to generate classification predictions during the learning phase. The average style features of each old domain in the style memory are projected onto the content features of the new domain image to synthesize the feature representation of the old domain image, and then input into the decoder to generate pseudo-labels of the old categories for the new domain image; based on the ground truth labels of the new categories and the pseudo-labels of the old categories, a loss function is designed to achieve cross-domain category knowledge accumulation. For new domain images, the same style randomization operation is applied to targets belonging to the same category, and the processing results of each category are combined to obtain a stylized image. By constraining the consistency loss between the stylized image and the original image in terms of content features, the quality of the synthesized features is improved.

2. The method according to claim 1, characterized in that: In the Each incremental learning phase will use the remote sensing images of the current domain. Input content encoders respectively and style encoder This generates corresponding depth features. , ,in Indicates spatial dimension, This refers to the channel dimension; the orthogonal constraint is achieved by minimizing the cosine similarity between content features and style features, specifically expressed as: in, This represents the training dataset for the current learning phase. It is cosine similarity. It is global average pooling.

3. The method according to claim 1, characterized in that: By driving volume The exponential moving average method is used to obtain style features of a specific domain. It is iteratively updated during training, and its update formula is as follows: in Indicates the first In each training iteration, from the input The extracted style features, and the domain-specific style features in the incremental learning phase, are denoted as... , The total number of iterations during the learning phase will be used to store this style feature in the style memory.

4. The method according to claim 1, characterized in that: A random fusion strategy is adopted, with random sampling probability. The choice between using domain-specific style features or style features extracted from the current batch for fusion is thus made during the training process. The calculation is as follows: in, For random selection or Indicates the first In each training iteration, from the input Extracted content features Indicates the first In each training iteration, from the input Extracted style features Represents domain-specific style characteristics, This indicates the fusion module, which consists of a convolutional layer, an AdaIN layer, and a ReLU activation layer.

5. The method according to claim 1, characterized in that: Adopting the first The extraction model generated during each learning stage produces pseudo-labels for the old categories: in This indicates that the extraction model in the previous learning stage is in the [number]th [stage]. The predicted classification probability of a new domain image under a domain-specific style. Represents domain-specific style characteristics; , , These are the content encoder, fusion module, and decoder from the previous learning stage.

6. The method according to claim 1, characterized in that: Truth labeling based on new categories and old category pseudo-labels Four loss functions were designed, the first of which is the new domain-new class learning loss function. The second old-new-class expansion loss function is used to supervise learning of new categories in a new domain. The third new-domain-old-class adaptation loss function is used to guide the model to extend knowledge of new categories to the old domain. The fourth old-domain-old-class preservation loss function is used for old-class adaptation on the new domain. Used to ensure that the current model does not forget the old category knowledge learned in the old domain.

7. The method according to claim 6, characterized in that: First new domain - new category learning loss function The function form is as follows: in, It is a decoder In receiving new domain features The resulting classification probability; Represents the model at pixel position Category The predicted probability, For remote sensing images, This represents the training dataset for the current learning phase; It is the expanded category space, containing all the old categories. and background categories ; Second old domain - new category extended loss function The function form is: in, It is by Based on intermediate variables The calculated value is used to solve the background offset problem. It is the decoder In receiving synthesized old domain features The resulting predictions; The third new domain-old category adaptation loss function The calculation process is as follows: First, the old model predictions are filtered using the following strategy to generate reliable pseudo-labels: Combining truth labels of new categories Optimize: Filtering function right Perform the screening. Finally... Calculated by the following formula: Similarly, These are the modified classification probabilities to address the background offset issue: ； The fourth old domain-old category preservation loss function The calculation form is as follows: It is by The calculation results are as follows: Based on the four loss functions mentioned above, the cross-domain category knowledge accumulation function is: = = This is an adjustable hyperparameter used to balance the weights of various loss terms.

8. The method according to claim 1, characterized in that: In the style randomization process, convolutional layers with random sampling weights are first used to enrich texture features, and then randomized style statistics are used to replace the original statistics, thereby achieving style diversification.

9. The method according to claim 1, characterized in that: The calculation of the content consistency constraint loss includes: inputting the content features of the stylized image, the content features of the original image, and the style features of the original image into the reconstruction module to generate a reconstructed image, and achieving content consistency constraint by minimizing the difference between the reconstructed image and the original image.

10. A class-domain incremental learning system for extracting features from the environment of power distribution networks, characterized in that: It includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the program instructions in the memory to execute the class-domain incremental learning method for extracting ground features in a power distribution network environment as described in any one of claims 1-9.

Images