An Incremental Image Classification Method Based on Imbalanced Modeling

Abstract

This invention discloses a class-incremental image classification method based on imbalanced modeling, applicable to image classification tasks under adverse conditions such as continuous increase in classes and drastic changes in distribution. First, a limited sample pool is used to preserve the distribution of the old classes, and the classification loss of the model is corrected based on the prior probability estimate of each class to obtain a more balanced classification boundary. Then, an old model mirror image is saved, and the old model mirror image is maintained and updated by the output classification model, while knowledge distillation is used to mitigate catastrophic forgetting. Finally, the regularization factor of the model training is dynamically adjusted to adapt to the continuous expansion of data classes. This invention can adaptively correct the parameters of the feature extractor to cope with imbalanced training data, and the improvement can be carried over to subsequent tasks through knowledge distillation, solving the problems of drastic decrease in image classification accuracy and new class bias in classification accuracy under the aforementioned adverse conditions.

Description

Technical Field

[0001] This invention relates to a class incremental image classification method based on imbalanced modeling, belonging to the fields of deep learning image classification and class incremental learning. It can perform data classification in streaming tasks where data categories dynamically increase and data distribution changes drastically, and can be applied to scenarios such as face recognition, autonomous driving, medical image analysis, behavior recognition, and image retrieval in situations with dynamically changing data distribution. Background Technology

[0002] Currently, state-of-the-art deep learning image classification algorithms have achieved levels comparable to or even surpassing human performance under fixed dataset training conditions. However, deep learning models still have some limitations; compared to humans, their ability to continuously learn knowledge is very limited. For example, when we feed a deep learning model that can already accurately classify some animal images with plant images, it often quickly forgets the previously learned animal knowledge while learning how to classify plants—a phenomenon known in academia as catastrophic forgetting. With the impact of the big data era, we urgently need to endow deep image classification models with the ability to continuously learn under dynamically changing data distributions.

[0003] This catastrophic forgetting phenomenon stems primarily from the high sensitivity of deep models to training data. Each training sample's learning process adjusts the model parameters, making the output more consistent with the new data distribution while simultaneously distancing it from the no longer accessible distribution of older data. Therefore, if old samples are not reviewed for a period, the model will catastrophically forget the old categories. A relatively simple solution is to save all previously learned images and merge them into the new data for joint training when a new category arrives. However, in today's big data era, saving all training data is extremely expensive. Furthermore, some continuous model expansion schemes also introduce greater storage pressure and exhibit poor generalization capabilities.

[0004] Meanwhile, other studies have pointed out that image classification models suffer from imbalanced and biased classification boundaries in continuous tasks. However, most of their solutions involve adjusting the biased output of the model after training or training a new set of parameters to correct this bias, without calibrating the parameters of the entire original model. Only by calibrating the parameters of the original model can this improvement be propagated across tasks, resulting in a considerable increase in accuracy. Furthermore, we hope that the deep learning image classification algorithms used can be well adapted and configured for different image classification models. Therefore, there is an urgent need for a class incremental image classification method based on imbalanced modeling. Summary of the Invention

[0005] The summary section of this application is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description section below. This summary section is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0006] To address the problems and shortcomings of existing technologies, this invention aims to provide a class incremental image classification method based on imbalanced modeling. For multi-class classification tasks in class incremental learning, this method uses a limited-capacity sample pool to retain historical data and statistically analyzes the prior probabilities of each class in the sample pool and new samples, using this to correct the originally imbalanced multi-class learning process. Simultaneously, we maintain a model mirror image and update it using the current model in each task to generate pseudo-labels for the current task. Knowledge distillation is employed to further mitigate catastrophic forgetting. Finally, we set dynamic parameter regularization factors for the classification model to adapt to the dynamically growing data volume in streaming tasks, thereby solving the problems mentioned in the background.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] This invention discloses a class incremental image classification method based on imbalance modeling, comprising the following steps:

[0009] Step 1: Obtain the initialized restricted sample pool M and its sample pool capacity constraint value m, and initialize the restricted sample pool M as an empty set;

[0010] Step 2: Randomly initialize the classification model Θ0 and the mirror image of the old classification model Θ0. old and regularization factor γ0;

[0011] Step 3: Obtain the total number of tasks T corresponding to the incremental tasks to perform the image classification task, and obtain the dataset D for the image classification task. i ;

[0012] Step 4, construct the total loss function And with the total loss function The objective function in dataset D I After training on ∪M until convergence, output the current task classification model Θ. I ;

[0013] Step 5, use the current task classification model Θ I Update the old classification model mirror Θ old ;

[0014] Step 6, for dataset D ISample M is performed, and the restricted sample pool M in the next task is updated sequentially until all image classification tasks are completed and the classification model is output.

[0015] Step 7: Input the image sample to be classified into the trained classification model, and obtain and output the image classification result.

[0016] Furthermore, when the image classification task performed in step 4 is the initial task, the cross-entropy classification loss function is used. The objective function is defined in the dataset D. I Training the classification model to convergence includes the following steps:

[0017] Step 4.1.1: Set the batch size b and initial learning rate η for mini-batch gradient descent;

[0018] Step 4.1.2, transfer the dataset D I Randomly split into multiple data parts of size b;

[0019] Step 4.1.3: Calculate the cross-entropy classification loss function for each data portion. and regularization term

[0020] Step 4.1.4, calculate the total loss function. Update the current task classification model Θ using gradient descent. i ;

[0021] Step 4.1.5: If the model converges, output the current task classification model Θ. I Otherwise, decay the learning rate η and jump to step 4.1.3 for the next round of training.

[0022] Furthermore, when the image classification task performed in step 4 is a non-initial task, a classification loss function calibrated with imbalanced class prior probabilities is used. Knowledge distillation loss function The objective function is defined in the dataset D. i Training the classification model on ∪M until convergence specifically includes the following steps:

[0023] Step 4.2.1: Set the batch size b and initial learning rate η for mini-batch gradient descent;

[0024] Step 4.2.2: Calculate the classification loss function used to reconcile the imbalanced class prior probabilities after calibration. and the knowledge distillation loss function The value of λ;

[0025] Step 4.2.3, transfer the dataset DI ∪M is randomly split into multiple dataset parts of size b;

[0026] Step 4.2.4: Calculate the classification loss function calibrated based on the imbalanced class prior probabilities for each dataset portion. Knowledge distillation loss function and regularization terms

[0027] Step 4.2.5 Calculate the total loss function Update the current task classification model Θ using gradient descent. i ;

[0028] Step 4.2.6: If the model converges, output the current task classification model Θ. i Otherwise, decay the learning rate η and jump to step 4.2.4 for the next round of training.

[0029] Furthermore, in step 6, the dataset D... i Sampling is performed on ∪M, and the restricted sample pool M in the next task is updated sequentially. The specific steps include:

[0030] In step 6, the dataset D i Sampling is performed on ∪M, and the restricted sample pool M in the next task is updated sequentially. The specific steps include:

[0031] Step 6.1: Initialize an empty set Ω, and obtain the maximum retention limit value n allocated to each category l by the restricted sample pool M;

[0032] Step 6.2: Use each category l to obtain the category feature center for each category.

[0033] Step 6.3: Calculate the sample χ closest to the class feature center using the maximum preservation limit value n. n and sample χ n Save to the set Ω;

[0034] Step 6.4: Update the restricted sample pool M using the set Ω, i.e., M = Ω.

[0035] Furthermore, the cross-entropy classification loss function Defined as,

[0036]

[0037] Among them, (X) j Y j Let C be the input and label of the j-th sample in the mini-batch data. allThe number of categories that have been seen. For the current task classification model Θ i The c-th dimension output.

[0038] Furthermore, the classification loss function after adjusting the imbalanced class prior probabilities. Defined as,

[0039]

[0040] Where, p c For the dataset D I The prior probability of class c estimated in ∪M is obtained by dividing the number of samples of class c by the total number of samples, and μ is a pre-set hyperparameter that controls the smoothness of the output distribution.

[0041] Furthermore, the knowledge distillation loss function Defined as,

[0042]

[0043] Among them, C old For the old category number, Mirror the old model Θ old The c-th dimension of the output is given by τ, which is a pre-set hyperparameter controlling the smoothness of the output distribution.

[0044] Furthermore, the regularization term Defined as,

[0045]

[0046] Among them, ||Θ i ||2 represents the sum of the L2 norms of all parameters in the model.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] This invention decomposes the problem of incremental learning into three parts: the training problem under imbalanced training sets, the problem of forgetting old category knowledge in streaming tasks, and the problem of dynamic data growth faced by the model in streaming tasks. Corresponding modules are proposed to address each of these problems. In particular, this invention has very low storage requirements; a single sample pool with a fixed capacity to store old category samples is sufficient to handle dynamic data changes and the continuous increase in categories. Furthermore, all parameters of the model trained on the imbalanced dataset are calibrated, resulting in a more balanced classification boundary. The gains from these parameter calibrations are then well inherited by the knowledge distillation module for the next task. Simultaneously, knowledge distillation technology is used to propagate the gains from the former along the task, mitigating catastrophic forgetting. Ultimately, this leads to a considerable improvement in classification accuracy, making it applicable to almost all deep learning image classification models. Attached Figure Description

[0049] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application.

[0050] In the attached diagram:

[0051] Figure 1 This is a flowchart illustrating the main steps in an embodiment of the present invention.

[0052] Figure 2 This is a flowchart illustrating the overall steps in an embodiment of the present invention.

[0053] Figure 3 This is a schematic diagram illustrating the steps of model training when performing the first image classification task in an embodiment of the present invention.

[0054] Figure 4 This is a schematic diagram illustrating the steps of model training when performing the second to the last image classification task in an embodiment of the present invention.

[0055] Figure 5 This is a schematic diagram illustrating the steps of sampling the dataset and updating the restricted sample pool in an embodiment of the present invention. Detailed Implementation

[0056] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0057] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0058] This invention discloses a class incremental image classification method based on imbalance modeling. The following will describe this disclosure in detail with reference to the accompanying drawings and embodiments.

[0059] Reference Figures 1 to 2 As shown, the present invention mainly includes the following steps:

[0060] Step 1: Obtain the initialized restricted sample pool M and its sample pool capacity constraint value m, and initialize the restricted sample pool M as an empty set;

[0061] Step 2: Randomly initialize the classification model Θ0 and the mirror image of the old classification model Θ0. old and regularization factor γ0;

[0062] Step 3: Obtain the total number of tasks T corresponding to the incremental tasks to perform the image classification task, and obtain the dataset D for the image classification task. i ;

[0063] Step 4, construct the total loss function and with the total loss function The objective function in dataset D i After training on ∪M until convergence, output the current task classification model Θ. i ;

[0064] Step 5, use the current task classification model Θ i Update the old classification model mirror Θ old ;

[0065] Step 6, for dataset D i Sample M is used to update the restricted sample pool M in the next task in turn, until all image classification tasks are completed and the classification model is output.

[0066] Step 7: Input the image sample to be classified into the trained classification model, and obtain and output the image classification result.

[0067] Specifically, we split a complete image dataset into T parts, with no overlap in the data categories of each part. Therefore, each part represents a completely new and unseen category for the model, and these T parts constitute T continuously arriving classification tasks. For the first initial task, we typically use traditional cross-entropy classification loss for training; for the other tasks, we employ class imbalance learning and knowledge distillation methods to construct the overall loss function. Let's consider the image classification dataset D obtained in step 3. i In practical applications, the dataset D may contain categories that the model has never seen before. This means that multiple new categories need to be trained on top of an already trained model. Therefore, it's necessary to maintain both a mirror image of the old model and a limited-capacity sample pool. i Sample from ∪M and update the restricted sample pool M in the next task accordingly. Continue until the number of image classification tasks performed is the same as the total number of tasks, then output the trained classification model.

[0068] More specifically, we use commercial data with 100 categories, including animals, plants, clothing, etc. We split this data into 10 parts to form 10 task datasets, each containing 10 new categories. We can use a 32-layer ResNet, ViT-M, or other appropriately sized image classification model, with a restricted sample pool M and a mirror image of the old model Θ. old Initially, all values ​​are empty. The initial value of the regularization factor γ0 is determined by the amount of data for a single task and the model; we can select the optimal value on the validation set. We will employ different model training procedures for the first task and other tasks, which will be explained in detail below. After each task, we need to update the classification model Θ for that current task. I Save as an old model mirror Θ old It is also necessary to use the current training dataset D i ∪M downsampling is used to obtain a restricted sample pool M for the next task, ensuring its capacity satisfies the constraint |M|≤m. It is important to note that the feature centers of each class in the sampling results should be as close as possible to the feature centers of all samples of that class in the training set.

[0069] In step 2, the initial random classification model Θ0 and the mirror image of the old classification model Θ are used. old And regularization factor γ0, and the initial classification model Θ0 is a mirror image of the initial old model Θ. old They are equal. In step 3, the total number of tasks corresponding to the incremental tasks is T, and the i-th image classification task is executed, i = 1, 2, ..., T. If i = 1, it is the first initial task. Then the cross-entropy classification loss function is used. The objective function is defined in dataset D. iTrain the classification model until convergence. Since there are no old classes in the first initial task, the training process differs from that of subsequent tasks, so only the traditional cross-entropy classification loss function needs to be used.

[0070] Reference Figure 3 As shown, when the image classification task performed in step 4 is the initial task, the cross-entropy classification loss function is used. The objective function is defined in dataset D. i Training the classification model until convergence includes the following steps:

[0071] Step 4.1.1: Set the batch size b and initial learning rate η for mini-batch gradient descent;

[0072] Step 4.1.2, transfer dataset D i Randomly split into multiple data parts of size b;

[0073] Step 4.1.3: Calculate the cross-entropy classification loss function for each data portion. and regularization term

[0074] Step 4.1.4, calculate the total loss function. Update the current task classification model Θ using gradient descent i ;

[0075] Step 4.1.5: If the model converges, output the current task classification model Θ. i Otherwise, decay the learning rate η and jump to step 4.1.3 for the next round of training.

[0076] Specifically, we use the mini-batch echeloned descent optimization algorithm to train the model. The batch size *b* and the initial learning rate *η* are both related to the training set and the model; *b* is typically set between 16 and 128, and *η* is typically set between 0.01 and 1. More specifically, a 32-layer ResNet network takes approximately 250 iterations to converge on our commercial dataset, with *b* = 32 and *η* = 0.1. After the 100th, 150th, and 200th iterations, the learning rate *η* is decayed to 1 / 10 of its original value. The total loss function in step 4.1.4... Represented as This is the cross-entropy classification loss function. With regularization term The summation. And the model parameters Θ updated using gradient descent. i Represented as In step 4.1.3, we use the traditional cross-entropy classification loss function, expressed as:

[0077]

[0078] Among them, (X) j Y j Let C be the input and label of the j-th sample in the mini-batch data. all C is the number of seen categories, in this case... all That is, 10. For the current task classification model Θ i The c-th dimension of the output. And the regularization term loss function in step 4.1.3. This is expressed as,

[0079]

[0080] Among them,}||Θ i ||2 represents the sum of the L2 norms of all parameters in the model, γ i This indicates that the regularization factor is updated to... By optimizing this regularization term This can effectively prevent model overfitting. Therefore, when the amount of data is small at the beginning of the task, we apply a larger regularization term, precisely because the model is more susceptible to overfitting at this time.

[0081] In the subsequent tasks following step 4, due to the existence of the old class, the model needs to simultaneously learn from both the old and new classes. When performing the image classification task, if i is not 1 and i < T, then the classification loss function is calibrated with the imbalanced class prior probability. and knowledge distillation loss function The objective function is defined in dataset D. i Train the classification model on ∪M until convergence.

[0082] Reference Figure 4 As shown, the classification loss function is calibrated using imbalanced class prior probabilities. Knowledge distillation loss function The objective function is defined in dataset D. i Training a classification model on ∪M until convergence includes the following steps:

[0083] Step 4.2.1: Set the batch size b and initial learning rate η for mini-batch gradient descent;

[0084] Step 4.2.2: Calculate the classification loss function used to reconcile the imbalanced class prior probabilities after calibration. and knowledge distillation loss function The value of λ;

[0085] Step 4.2.3, transfer dataset D I ∪M is randomly split into multiple dataset parts of size b;

[0086] Step 4.2.4: Calculate the classification loss function calibrated based on the imbalanced class prior probabilities for each dataset portion. Knowledge distillation loss function and regularization terms

[0087] Step 4.2.5 Calculate the total loss function Update the current task classification model Θ using gradient descent i ;

[0088] Step 4.2.6: If the model converges, output the current task classification model Θ. i Otherwise, decay the learning rate η and jump to step 4.2.4 for the next round of training.

[0089] Specifically, in our commercial dataset instance, the batch size *b*, initial learning rate *η*, optimization algorithm, and learning rate decay rate for subsequent tasks remain consistent with those set in the first initial task, because the amount of new data for each task is exactly the same. Of course, when the amount of data for each task changes, we need to adjust the training parameters for each task accordingly. Starting with the second task, we add a step to calculate *λ*, which is used to coordinate the classification loss function after calibrating the imbalanced class prior probabilities, as mentioned later. And the knowledge distillation loss function to prevent catastrophic forgetting because More focus is placed on learning new categories. Greater emphasis is placed on reviewing previously encountered categories, and this parameter increases progressively as the task progresses. This is because the number of previously encountered categories increases with each task, making review increasingly important. The total loss function in step 4.2.5... Represented as The model parameters Θ are updated using gradient descent. i Also expressed as The classification loss function after imbalanced class prior probabilities calibration Calculated using the following formula:

[0090]

[0091] Where, p c For the dataset D i The prior probability of class c estimated in ∪M is obtained by dividing the number of samples of class c by the total number of samples. μ is a pre-set hyperparameter controlling the smoothness of the output distribution. μ is generally set between 1 and 2; here, it is set to 1.5 on our commercial dataset. Knowledge distillation loss function. Calculated using the following formula:

[0092]

[0093] Among them, C old This represents the number of old categories (i.e., the total number of categories seen in the previous task). Mirroring the old model Θ old The c-th dimension of the output is given, where τ is a pre-set hyperparameter controlling the smoothness of the output distribution. τ is typically set between 2 and 8; here, it is set to 4 on our commercial dataset.

[0094] As described in steps 5 and 6, after each image classification task is completed, we need to update the current task classification model Θ. i Save as an old model mirror Θ old , that is, let Θ old =Θ i And it also needs to be in the current training dataset D. i ∪M downsampling is used to obtain a restricted sample pool M for the next task. It is important to note that the class center of each class l in the sampling result should be as close as possible to the class feature centers of all samples of that class in the training set.

[0095] Reference Figure 5 As shown, for dataset D i Sampling is performed on ∪M, and the restricted sample pool M in the next task is updated sequentially. The specific steps include:

[0096] Step 6.1: Initialize an empty set Ω and obtain the maximum storage limit value n allocated to each class l by the restricted sample pool M;

[0097] Step 6.2, use each category l to obtain the category feature center for each category.

[0098] Step 6.3, using the maximum preservation limit value n, calculate the sample χ closest to the class feature center. n and sample χ n Save to the set Ω;

[0099] Step 6.4: Update the restricted sample pool M using the set Ω, i.e., M = Ω.

[0100] Specifically, for each category l, l = 1, 2, ..., C all Maximum storage limit n, We aim to maintain a balanced, constrained sample pool M, meaning that the storage amount for each class should be nearly consistent. Therefore, we impose a maximum storage limit value n on the storage amount for each class. Furthermore, we aim to ensure that the samples stored in the sample pool can reasonably represent their respective categories. Therefore, we want the feature centers of these samples to be as consistent as possible with the feature centers of all samples of that category. Therefore, we will... In each iteration of the round, a value is calculated that makes the saved sample center closest to the class center. Sample χ n And include it in the next round's sample pool M. Sample χ n Represented as

[0101] More specifically, in our commercial dataset, each task has approximately 50,000 samples, while our sample pool limit m=2000 is much smaller than that. When the last task is executed, less than 23 samples are saved for each old class, making the storage overhead very inexpensive.

[0102] This disclosure endows image classification models with the ability to continuously expand categories. As commercial data continues to arrive, the parameters of the entire classification model are calibrated through a reliable description of this bias using prior probabilities for each category, without incurring excessive sample storage burden. Furthermore, this disclosure does not involve changes to the model structure and is applicable to almost all deep learning image classification models. In today's world of ever-growing commercial data volumes, it can effectively handle dynamic data changes and the continuous increase in categories.

[0103] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A class incremental image classification method based on imbalanced modeling, characterized in that, Includes the following steps: Step 1: Obtain the initialized restricted sample pool and its sample pool capacity constraint value and the restricted sample pool Initialize to an empty set; Step 2, randomly initialize the classification model Old classification model mirror and regularization factor ; Step 3: Obtain the total number of tasks corresponding to the incremental tasks. To perform an image classification task and obtain the image classification dataset. ; Step 4, construct the total loss function and with the total loss function For the objective function in the dataset After training to convergence, output the classification model for the current task. ; Step 5, use the current task classification model. Update the old classification model image ; Step 6, process the dataset Sampling is performed, and the restricted sample pool in the next task is updated sequentially. The classification model is output after all image classification tasks have been completed. Step 7: Input the image sample to be classified into the trained classification model, and obtain and output the image classification result; When the image classification task performed in step 4 is not the initial task, the classification loss function calibrated with imbalanced class prior probabilities is used. Knowledge distillation loss function The objective function is defined in the dataset. The process of training the classification model to convergence includes the following steps: Step 4.2.1: Set the batch size for mini-batch gradient descent. and initial learning rate ; Step 4.2.2: Calculate the classification loss function used to reconcile the imbalanced class prior probabilities after calibration. and the knowledge distillation loss function of value; Step 4.2.3, transfer the dataset Randomly split into multiple sizes The dataset portion; Step 4.2.4: Calculate the classification loss function calibrated based on the imbalanced class prior probabilities for each dataset portion. Knowledge distillation loss function and regularization terms ; Step 4.2.5, calculate the total loss function. And update the current task classification model using gradient descent. ; Step 4.2.6: If the model converges, output the current task classification model. Otherwise, decay the learning rate. Then proceed to step 4.2.4 for the next round of training; The classification loss function after adjusting the imbalanced class prior probabilities Defined as, in, For the first in a small batch of data The input and labels of each sample, The number of categories that have been seen. For the current task classification model The Dimensional output, For the dataset The estimated category The prior probability is determined by the category. The sample size is obtained by dividing the total sample size. The hyperparameters that are preset to control the smoothness of the output distribution; Knowledge distillation loss function Defined as, in, For the old category number, Mirror the old classification model The Dimensional output, The hyperparameters are preset to control the smoothness of the output distribution.

2. The class incremental image classification method based on imbalanced modeling according to claim 1, characterized in that; When the image classification task performed in step 4 is the initial task, the cross-entropy classification loss function is used. The objective function is defined in the dataset. Training the classification model to convergence includes the following steps: Step 4.1.1: Set the batch size for mini-batch gradient descent. and initial learning rate ; Step 4.1.2, transfer the dataset Randomly split into multiple sizes The data section; Step 4.1.3: Calculate the cross-entropy classification loss function for each data portion. and regularization term ; Step 4.1.4, calculate the total loss function. And update the current task classification model using gradient descent. ; Step 4.1.5: If the model converges, output the current task classification model. Otherwise, decay the learning rate. Proceed to step 4.1.3 for the next round of training.

3. The class incremental image classification method based on imbalanced modeling according to claim 1, characterized in that, Step 6 involves processing the dataset. Sampling is performed, and the restricted sample pool in the next task is updated sequentially. The specific steps include: Step 6.1, initialize an empty set Obtain the restricted sample pool For each category Maximum storage limit allocated ; Step 6.2, using each of the categories Execute to obtain the category feature center for each category ; Step 6.3, use the maximum storage limit value. Perform calculation of the sample closest to the class feature center and the sample Save to the set middle; Step 6.4, using the set Update the restricted sample pool ,Right now .

4. The class incremental image classification method based on imbalanced modeling according to claim 2, characterized in that: The cross-entropy classification loss function Defined as, in, For the first in a small batch of data The input and labels of each sample, The number of categories that have been seen. For the current task classification model The Dimensional output.

5. The class incremental image classification method based on imbalanced modeling according to claim 3, characterized in that: The regularization term Defined as, in, This represents the sum of the L2 norms of all parameters in the model.

Images