Non-paradigm-based efficient knowledge perception incremental learning model establishment method and device

Abstract

The invention discloses a non-paradigm-based efficient knowledge perception incremental learning model establishment method. The method comprises the following steps: preprocessing an image in a training data set; constructing a class incremental learning efficient knowledge representation module and an image classification model, wherein the class incremental learning efficient knowledge representation module comprises a confidence prototype sampling module used for generating robust and diversified class specific prototype distribution and an adaptive memory regularization module used for regularizing increment classifier output; the image classification model comprises a feature extractor used for extracting features of an input image, and a prior guide increment classifier module which decouples new and old weights, freezes the weights, and promotes an old classifier to adapt to a new feature space while reducing new and old knowledge interference by extending a fine tuning component; the feature extractor is used for extracting features of an input image. According to the method, a class incremental learning task can be efficiently completed, and an image classification result is better than that of an existing class incremental learning method.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, and in particular to a high-efficiency knowledge-aware incremental learning model establishment method and device based on no examples. BACKGROUND

[0002] Incremental learning plays a crucial role in the field of computer vision, especially in the open world where data continues to grow, such as medical diagnosis, autonomous driving, and face recognition. Incremental learning aims to enable learning systems to simulate the human brain, continuously learning new knowledge from new samples, and retaining most of the previously learned knowledge during this process. It allows learning systems to continuously learn from new data rather than starting from scratch, thereby achieving accumulation and updating of knowledge. However, under this learning paradigm, incremental learning suffers from a serious "catastrophic forgetting" problem. Therefore, how to effectively resist catastrophic forgetting becomes a core problem in building an incremental learning model.

[0003] Class-incremental learning is a specific form of incremental learning, which aims to continuously learn new classes under continuous data flow, and the model needs to be able to distinguish all learned classes without task labels throughout the learning process. Today, the mainstream class-incremental learning method is the replay-based incremental learning method. This method often stores some representative old task samples when learning new knowledge. However, due to the need to continuously store old class samples, it may cause storage limitations or privacy security problems. Therefore, research methods for class-incremental learning based on no examples are emerging. This type of method does not need to save previous task samples during the incremental stage, and through a series of strategies, it deals with the problem of catastrophic forgetting. In particular, the prototype-based method has achieved remarkable performance under this paradigm. However, it still has the following problems: on the one hand, existing methods assume that the incremental learning model can fully adapt to the training samples of each stage, and the prototype derived from the indiscriminate sampling instances can represent each class. On the other hand, the incremental classifier usually adjusts to the new task by fine-tuning the parameters of the previous stage, while trying to retain previous knowledge. However, due to changes in data distribution, it may cause feature space drift and rewriting of old knowledge by the classifier. Therefore, this type of method still suffers from some forgetting problems. SUMMARY

[0004] The purpose of the present application is to provide a high-efficiency knowledge-aware incremental learning model establishment method and device based on no examples, which can efficiently complete the class-incremental learning task, and the image classification result is better than that of the existing class-incremental learning method.

[0005] To achieve the above technical purpose, the technical scheme adopted by the present application is as follows:

[0006] In a first aspect, the present application discloses a method for establishing a high-efficiency knowledge-aware incremental learning model based on no examples, which comprises the following steps:

[0007] S1: dividing the pictures in a data set of a plurality of images to be classified into a series of tasks according to the categories to which the pictures belong, taking the first task as a base task, and taking the remaining categories as incremental categories and dividing them according to the number of tasks, and then performing subsequent training;

[0008] S2: pre-processing the pictures in the training data set;

[0009] S3: constructing a high-efficiency knowledge-aware incremental learning model based on no examples, including a confident prototype sampling module, an adaptive memory regularization module and a prior guided incremental classifier module;

[0010] The confident prototype sampling module is used to screen a confident sample set from all sample network probabilities corresponding to each category to reduce the influence of incomplete fitting samples on the representation of old knowledge, generate a robust and diversified category prototype distribution, and dynamically reshape the old category prototype distribution to adapt to the changing feature space; the adaptive memory regularization module regularizes the output of the incremental classifier by saving old memories, and updates the old memory unit using the gradient momentum of the incremental classifier through a gradient evaluation strategy to make the old memories change with the new feature space; the prior guided incremental classifier module decouples the classifier, so that the model updates the new class classifier online while learning low-rank representation to ensure that the old class classifier adapts to the new feature space, and reduces knowledge forgetting caused by old parameter fine-tuning through historical memory guidance;

[0011] S4: sequentially obtaining the pre-processed training data, and training the high-efficiency knowledge-aware incremental learning model based on no examples;

[0012] S5: saving the high-efficiency knowledge-aware incremental learning model based on no examples corresponding to the current task as the old network of the next incremental stage;

[0013] S6: testing the saved model using the pre-processed test data set to obtain the average classification accuracy of the model for all known categories of images;

[0014] S7: returning to step S4 until the training of all tasks is completed, obtaining the trained high-efficiency knowledge-aware incremental learning model based on no examples, and using the trained high-efficiency knowledge-aware incremental learning model for image classification.

[0015] Further, in step S2, the pre-processing process of the images in the training data set comprises the following steps:

[0016] Step S2-1: randomly cropping the pictures to a predetermined size;

[0017] Step S2-2: The cropped picture is randomly horizontally flipped with a preset probability or randomly adjusted in brightness with a preset offset amplitude, and data is augmented.

[0018] Further, the confidence prototype sampling module generates robust and diverse class prototype distributions, and a process of dynamically reshaping an old class prototype distribution includes the following steps:

[0019] After the training of each stage is completed, the current stage training samples are again sent into the network to obtain the prediction probability output of each sample, and the samples in the training samples whose prediction probability output for each class is greater than a predefined threshold δ are taken as confident samples, and the confident sample set is calculated according to the following formula:

[0020]

[0021] In the formula, represents the sample corresponding to the t-th task, represents the sample after the feature extractor obtains the feature, θ t is the parameter of the feature extractor , represents the index of the maximum value, Y(·) represents the efficient knowledge-aware incremental learning model based on no example, I{·,·} represents the indicator function, v p represents the p-th class in the t-th task, is the total number of classes in the t-th task, Y t i represents the real label corresponding to the sample, j represents the predicted label after the network, N t represents the number of samples;

[0022] And the mean and covariance matrix of the confident sample set are calculated according to the following formula to form a class-specific multivariate normal distribution:

[0023]

[0024] In the formula, represents the prototype feature of class v sampled from the multivariate normal distribution p , represents the mean of the confident sample feature corresponding to class v p , represents the covariance matrix of the confident sample feature corresponding to class v p , represents the number of confident samples in class v p , is the i-th confident sample feature belonging to the category v p is the i-th confident sample feature belonging to the category v

[0025] At the beginning of the next phase, the multivariate normal distribution of each category is sampled to obtain the prototype feature of the old category The decision boundary of the prototype of the old category is dynamically reshaped by randomly weighting the prototype and the new category feature to expand the decision boundary of the prototype of the old category.

[0026] Further, the adaptive memory regularization module saves the classifier weights of the current t-1 task as old memory units in the memory pool after each task is completed During the learning of the new task t, for the forward propagation process, the increment classifier is optimized by the cross-entropy loss During backpropagation, the gradient p of the increment classifier is used to optimize the classifier. in Each old memory unit is momentum updated, and the dynamic implicit memory loss L dhm (ξ) is used to regularize the output of the classifier with the aid of new knowledge.

[0027] Further, the calculation method of the dynamic implicit memory loss L dhm (ξ) is as follows:

[0028]

[0029] In the formula, ζ represents a scaling factor, h is the batch size of the current task t phase, is the old memory unit selected from the memory pool, φ t-i represents the parameters of the old memory unit, f new represents the feature obtained by the sample of the new task through the feature extractor , P(ν p ) = (p(1,ν p |f new ), p(2,ν p |f new ),..., p(η t-i ,ν p |f new ) T represents the output of the fine-tuned increment classifier module, and the element calculation formula of p(i,v p |f new ) is as follows:

[0030]

[0031] In the formula, is the fine-tuned increment classifier of the t-th task phase, η t-i is the corresponding classifier​ dimension.

[0032] Further, the prior guided incremental classifier module comprises a fine-tuning incremental classifier and a self-supervised augmented classifier:

[0033] Wherein, at the tth task stage, the fine-tuning incremental classifier inherits the classifier weights from the previous stage and copies them twice to obtain two classifiers equivalent to the old task category dimension. For the two classifiers, two fine-tuning components are extended to adapt to new categories, respectively defined as and A projection function Proj(·) is added to the classifier to project the fine-tuning result to the new task dimension, and finally the outputs of the two classifiers are combined to obtain The calculation formula is as follows:

[0034]

[0035] In the formula, Concat(·,·) represents the aggregation of two input features along the channel dimension, f is the feature obtained by the sample passing through the feature extractor , and LoRA(·) is a fine-tuning component composed of two linear layers.

[0036] The self-supervised augmented classifier is used to obtain output probabilities by putting augmented images into it, and to enhance the generalization ability of the model by self-supervised contrastive learning with the output probabilities of the fine-tuning incremental classifier.

[0037] Step S4 comprises:

[0038] Step S4-1: setting training parameters, if the current task is the base task, randomly initializing the efficient knowledge-aware incremental learning model based on no examples, if it is an incremental task, directly inheriting the network parameters obtained by training the previous task to initialize;

[0039] Step S4-2: rotating the ith training image D t of the tth task training data set D t,i by 4 angles (0°, 90°, 180°, 270°) to obtain the rotated augmented image and calculating the cross-entropy loss and the self-supervised loss and optimizing the model; if the current task belongs to an incremental task, continue to execute step S4-3, otherwise, go to step S4-5;

[0040] Step S4-3: rotating the ith training image D The knowledge distillation loss is calculated by feeding the data into the network trained for the current task and the frozen network from the previous task. The prior-guided fine-tuning incremental classifier module is executed to calculate the class separation loss.

[0041] Step S4-4: By from The old category prototype features ξ are obtained by sampling from the distribution, and the dynamic implicit loss L is calculated through adaptive memory regularization. dhm (ξ);

[0042] Step S4-5: If the current task belongs to the basic task, use the cross-entropy loss calculated in step S4-2. and self-monitoring loss The network parameters are optimized, and the loss is calculated as follows:

[0043]

[0044] If the current task is an incremental task, combine cross-entropy loss. Self-monitoring loss Knowledge distillation loss Class separation loss and dynamic implicit loss L dhm (ξ), the network parameters are optimized, and the loss is calculated as follows:

[0045]

[0046] In the formula, L is the total loss of the model, and λ1 and λ2 are hyperparameters that control the importance of the loss;

[0047] Step S4-6: After the current stage ends, establish a confidence prototype sampling method for each category. The distribution updates the prototype library and saves the current stage incremental classifier weights to update the memory pool. in t represents the current number of tasks, and h represents the number of old memory units in the memory pool.

[0048] Furthermore, the self-monitored loss The specific calculation method is as follows:

[0049]

[0050] In the formula, KL(·) is the KL divergence loss, P represents the probability distribution of the classifier output, and f is the probability distribution of the current stage D. t The samples in the sample are processed by the feature extractor The obtained features Where f aug The sample is rotated four angles through the feature extractor the obtained feature, is a self-supervised augmented classifier parameterized by K.

[0051] Further, the calculation method of the knowledge distillation loss is specifically as follows:

[0052]

[0053] wherein f old is the output feature of the sample in the current stage obtained by the old network feature extractor , ||·||2 is an L2 distance loss, and f is the feature of the sample in the current stage obtained by the feature extractor ;

[0054] The calculation method of the knowledge distillation loss is specifically as follows:

[0055]

[0056] wherein d(·) is the Euclidean distance between two samples, a is an anchor sample, p is a positive sample, i.e., a sample with the same label as the anchor sample, n is a negative sample, i.e., a sample with a different label from the anchor sample, and m is a boundary parameter for forcing the desired separation between the positive and negative samples.

[0057] In a second aspect, the present application discloses a no-example-based efficient knowledge-aware incremental learning model device, which comprises a processor and a computer program stored on a memory and capable of running on the processor, and the processor implements the method as described above when executing the program.

[0058] Compared with the prior art, the present application has the following beneficial effects:

[0059] First, the no-example-based efficient knowledge-aware incremental learning model establishment method and device of the present application use a confidence prototype sampling method to reduce the influence of incomplete fitting samples, and the evaluation index reaches the optimal result in the prior art.

[0060] Second, the no-example-based efficient knowledge-aware incremental learning model establishment method and device of the present application design an adaptive memory regularization method, which gives the old memory unit a gradient through gradient evaluation, thereby providing the old knowledge with the most valuable information for new tasks, and adaptively regularizing the output of the incremental classifier through historical knowledge under the assistance of the above, thereby effectively reducing the catastrophic forgetting problem.

[0061] Third, the prior guided fine-tuning incremental classifier module of the no-example-based efficient knowledge perception incremental learning model establishment method and device of the present application is proposed, which decomposes the incremental classifier, reduces the interference between the new and old weights through the guidance of the fine-tuning component, and promotes the adaptation of the old classifier in the new feature space. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 is a flowchart of the no-example-based efficient knowledge perception incremental learning model establishment method provided by the present application;

[0063] Figure 2 is a structural diagram of the no-example-based efficient knowledge perception incremental learning model of the present application. DETAILED DESCRIPTION

[0064] The embodiments of the present application are further described in detail below with reference to the accompanying drawings.

[0065] The present embodiment specifically provides a no-example-based efficient knowledge perception incremental learning model establishment method, as shown in Figure 1 and Figure 2 , specifically comprising the following steps:

[0066] Step S1: dividing a data set containing a plurality of images to be counted into a training data set, a validation data set and a test data set, and determining the learning order of different categories of images in the data set.

[0067] The present embodiment selects Cifar-100, Tiny-ImageNet and ImageNet-100 data sets for class incremental learning. Among them, the first 40 (when the number of tasks is 20) or 50 categories are selected from the data set as the base categories of the first task, and the remaining categories are divided equally according to the number of tasks 5, 10 and 20. First, the base categories are trained, and the remaining tasks are used as incremental tasks for subsequent training.

[0068] Step S2: pre-processing the images in the training data set; wherein the pre-processing method is specifically as follows:

[0069] Step S2-1: randomly cropping the image to a preset size; in the present example, the preset sizes according to the three different data sets are 32*32, 64*64 and 224*224.

[0070] Step S2-2: data augmentation is performed on the cropped image. The method of data augmentation is: S2-2-1: the image is randomly horizontally flipped with a preset probability; S2-2-2: the image is randomly adjusted in brightness with a preset offset amplitude. In the embodiment, the preset probability of random horizontal flipping and random adjustment of image brightness is 0.5 and 0.24705882352941178, respectively.

[0071] Step S3: constructing a class-incremental learning efficient knowledge representation module and an image classification model, wherein the class-incremental learning efficient knowledge representation module comprises a confidence prototype sampling module and a self-adaptive memory regularization module, and the image classification model comprises a feature extractor and a prior-guided fine-tuning incremental classifier module.

[0072] Step S3-1: the specific operation of the confidence prototype sampling module comprises:

[0073] After the training of each stage is completed, the prediction probability output of each sample is obtained by sending the training sample of the current stage into the network again, and the sample in the training sample whose prediction probability output is greater than the predefined threshold δ for each class is taken as a confident sample. The confident sample set is calculated according to the following formula:

[0074]

[0075] In the formula, represents the sample corresponding to the t-th task, represents the feature obtained by the sample passing through the feature extractor θ t is the parameter of the feature extractor , represents the index of the maximum value, γ(·) represents the efficient knowledge perception incremental learning model based on no example, I{·,·} represents the indicator function, v p represents the p-th class in the t-th task, is the total number of classes in the t-th task, Y t i represents the real label corresponding to the sample, j represents the predicted label after passing through the network, N t represents the number of samples;

[0076] And the mean and covariance matrix of the confident sample set are calculated according to the following formula to form a class-specific multivariate normal distribution:

[0077]

[0078]

[0079] In the formula, Indicates a multivariate normal distribution The category v obtained from mid-sampling p Prototype features, Represents category v p The mean of the corresponding confident sample features, Represents category v p The covariance matrix of the corresponding confident sample features, Represents category v p The number of self-confident samples in China For belonging to category v p The features of the i-th confident sample;

[0080] At the start of the next phase, the old class prototype features are obtained by sampling the multivariate normal distribution of each class. Then, the old class distribution is dynamically reshaped by randomly weighting the prototypes with the new class features to expand the decision boundary of the old class prototypes.

[0081] Step S3-2: The specific operations of the adaptive memory regularization module include:

[0082] After each task is completed, the classifier weights of the current task t-1 are saved as old memory units in the memory pool. When learning a new task t, for the forward propagation process, cross-entropy loss is used. The incremental classifier is optimized, and its gradient p is utilized during backpropagation. in Momentum updates are performed on each old memory unit, and dynamic implicit memory loss L is applied. dhm (ξ) Regularize the classifier output with the help of new knowledge.

[0083] Among them, the dynamic implicit memory loss L in step S3-2 dhm The specific method for calculating (ξ) is as follows:

[0084]

[0085] In the formula, ζ represents the scaling factor, h is the batch size of the current task at stage t, and φ t-i The classifier weights for task phase ti. φ represents the old memory unit selected from the memory pool. t-i The parameter f represents the old memory unit. new This indicates that the samples for the new task have been processed by the feature extractor. The features obtained,

[0086] P(v p )=(p(1,v p |f new ),p(2,ν p |fnew ),...,p(η t-i ,v p |f new )) T denotes the output of the fine-tuning incremental classifier module, and the elements in p(i,v p |f new ) are calculated as follows:

[0087]

[0088] wherein, is the fine-tuning incremental classifier at the t-th stage, η t-i is the dimension of the corresponding classifier .

[0089] Step S3-3: The prior-guided incremental classifier module specifically comprises:

[0090] The fine-tuning incremental classifier is used to decouple the classifier weights and reduce the interference of new and old classification weights; through the guidance of the fine-tuning component, the old classifier is promoted to adapt to the new feature space.

[0091] The self-supervised augmented classifier is used to obtain output probabilities by placing augmented images therein, and to enhance the generalization ability of the model through self-supervised contrastive learning with the output probabilities of the fine-tuning incremental classifier.

[0092] Wherein, the method of the fine-tuning incremental classifier in step S3-3 is specifically:

[0093] At the t-th task stage, the fine-tuning incremental classifier inherits the classifier weights from the previous stage and copies them twice to obtain two classifiers equivalent to the old task class dimension. Two fine-tuning components are extended to adapt to new classes, respectively defined as and A projection function Proj(·) is added to the classifier to project the fine-tuning result to the new task dimension, and finally the outputs of the two classifiers are combined to obtain The calculation formula is as follows:

[0094]

[0095] wherein, Concat(·,·) represents the aggregation of two input features along the channel dimension, f is the feature obtained by the sample after the feature extractor , and LoRA(·) is the fine-tuning component composed of two linear layers.

[0096] Step S4: sequentially obtain the preprocessed training data, and train the efficient knowledge-aware few-shot incremental learning model. The training method specifically includes:

[0097] Step S4-1: set the training parameters, if the current task is the base task, randomly initialize the efficient knowledge-aware few-shot incremental learning model, if it is an incremental task, directly inherit the network parameters obtained by the last stage training to initialize;

[0098] Step S4-2: input the i-th training image D t of the t-th task training data set D t,i Rotate 4 angles (0°, 90°, 180°, 270°) to obtain the rotated augmented image Input the model to calculate the cross-entropy loss And the self-supervised loss Optimize the model, if the current task belongs to the incremental task, continue to execute step S4-3.

[0099] Among them, the calculation method of the self-supervised loss In step S4-2 specifically includes:

[0100]

[0101] In the formula, KL(·) is the KL divergence, P represents the probability distribution output by the classifier, f is the feature obtained by the sample in the current stage D t Through the feature extractor , Where is a self-supervised augmented classifier parameterized by κ, f aug is the feature obtained by rotating the sample by four angles through the feature extractor .

[0102] Step S4-3: put Into the network trained in the current stage and the frozen last task network, calculate the knowledge distillation loss Execute the prior-guided fine-tuning incremental classifier module to calculate the class separation loss Among them, the calculation method of the knowledge distillation loss And the class separation loss Specifically includes:

[0103] Step S4-3-1: calculate the knowledge distillation loss

[0104]

[0105] In the formula, f old is the feature obtained by the sample in the current stage samples in the old network feature extractor the output feature, ||·||2 is L2 distance loss, f is samples in the feature extractor the obtained feature;

[0106] Step S4-3-2: calculate the class separation loss according to the following formula

[0107]

[0108] In the formula, d(·) is the Euclidean distance between two samples, a is the anchor sample, p is the positive sample, i.e., the sample with the same label as the anchor sample, n is the negative sample, i.e., the sample with a different label from the anchor sample, and m is a boundary parameter for forcing the expected separation between positive and negative samples.

[0109] Step S4-4: if the current stage is not the base stage, obtain the old class prototype feature ξ by sampling from , calculate the dynamic implicit loss L dhm (ξ) through adaptive memory regularization.

[0110] Step S4-5: if the current stage is the base task, optimize the network parameters using the loss calculated in step S4-2, and the loss is calculated as follows:

[0111]

[0112] If the current task is an incremental task, optimize the network parameters in combination with the loss calculated in steps S4-3 to S4-4, and the loss is calculated as follows:

[0113]

[0114] In the formula, L is the total loss of the model, λ1 and λ2 are hyperparameters for controlling the importance of the loss.

[0115] Step S4-6: after the current stage is completed, update the prototype library for each class by establishing distribution through the confidence prototype sampling method, and save the incremental classifier weight update memory pool of the current stage wherein t represents the current number of tasks, and h represents the number of old memory units in the memory pool.

[0116] Step S5: save a copy independent of the current model by deep copying, as the old network of the next incremental stage, and then put the training data into the trained network.

[0117] The method of saving a copy independent of the current model by deep copy is as follows:

[0118] A new network model is created, and the weight parameters of the current model are copied to the newly created network model recursively, so there is no data sharing between the newly created network model and the current model, which is the old network of the next incremental stage.

[0119] Step S6: Test the saved model using the test data set to obtain the average accuracy of the model for all known categories of images;

[0120] Step S7: Return to step S4 until the training of all tasks is completed, and a trained efficient knowledge-aware incremental learning model is obtained.

[0121] The embodiment also provides a device for establishing an efficient knowledge-aware incremental learning model based on no examples, which comprises a processor and a computer program stored on a memory and executable on the processor, and the processor implements the above method when executing the program.

[0122] Although the preferred embodiments of the present application have been described, those skilled in the art can make further changes and modifications to the embodiments once they know the basic inventive concept. Therefore, the appended claims are intended to be interpreted as including all the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0123] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A high-efficiency knowledge perception incremental learning model based on no-example, characterized in that, The method comprises the following steps: S1: dividing the pictures in the data set of a plurality of images to be classified into a series of tasks according to the categories to which the pictures belong, taking the first task as a base task, and taking the remaining categories as incremental categories and dividing them according to the number of tasks, and then performing subsequent training; S2: preprocessing the pictures in the training data set; S3: constructing an efficient knowledge-aware incremental learning model based on no examples, including a confident prototype sampling module, an adaptive memory regularization module, and a prior guided incremental classifier module; The confident prototype sampling module is used to screen a confident sample set according to the network probability of all samples corresponding to each category to reduce the influence of incomplete fitting samples on the representation of old knowledge, generate a robust and diversified category prototype distribution, and dynamically reshape the old category prototype distribution to adapt to the changing feature space; the adaptive memory regularization module regularizes the output of the incremental classifier by saving old memories, and uses the gradient momentum of the incremental classifier to dynamically update the old memory unit through a gradient evaluation strategy to make the old memories change with the new feature space; the prior guided incremental classifier module decouples the classifier, so that the model updates the new class classifier online, learns low-rank representation to ensure that the old class classifier adapts to the new feature space, and reduces knowledge forgetting caused by old parameter fine-tuning through historical memory guidance; S4: sequentially obtaining the preprocessed training data, and training the efficient knowledge-aware incremental learning model based on no examples; S5: saving the efficient knowledge-aware incremental learning model based on no examples corresponding to the current task as the old network of the next incremental stage; S6: testing the saved model using the preprocessed test data set to obtain the average classification accuracy of the model for all known category images; S7: returning to step S4 until the training of all tasks is completed, obtaining the trained efficient knowledge-aware incremental learning model based on no examples, and using the trained efficient knowledge-aware incremental learning model for image classification. 2.The case-based efficient knowledge perception incremental learning model building method according to claim 1, characterized in that, In step S2, the process of preprocessing the pictures in the training data set comprises the following steps: Step S2-1: randomly cropping the pictures to a preset size; Step S2-2: randomly flipping the cropped pictures horizontally with a preset probability or randomly adjusting the brightness of the pictures with a preset offset amplitude to augment the data. 3.The case-based efficient knowledge perception incremental learning model building method according to claim 1, characterized in that, The process of generating a robust and diversified category prototype distribution by dynamically reshaping the old category prototype distribution of the confident prototype sampling module comprises the following steps: After the training of each stage is completed, the training samples of the current stage are again input into the network to obtain the prediction probability output of each sample, and the samples in the training samples whose prediction probability output for each category is greater than a predefined threshold δ are taken as confident samples, and the confident sample set is calculated according to the following formula: wherein, denotes the sample corresponding to the t-th task, denotes the sample passing through the feature extractor the obtained feature, t is the parameter of the feature extractor denotes the index of the maximum value, γ(·) denotes the efficient knowledge-aware incremental learning model based on no example, I{·,·} denotes the indicator function, v p denotes the p-th class in the t-th task, is the total number of classes in the t-th task, denotes the real label corresponding to the sample, j denotes the predicted label after passing through the network, N t denotes the number of samples;​ And the mean and covariance matrix of the confident sample set are calculated to form a category-specific multivariate normal distribution according to the following formula: In the formula, Indicates a multivariate normal distribution The category v obtained from mid-sampling p Prototype features, Represents category v p The mean of the corresponding confident sample features, Represents category v p The covariance matrix of the corresponding confident sample features, Represents category v p The number of self-confident samples in China For belonging to category v p The features of the i-th confident sample; At the beginning of the next phase, a multivariate normal distribution is performed for each class Sampling is performed to obtain old class prototype features The old class distribution is then dynamically reshaped to expand the decision boundary of the old class prototype by randomly weighting the prototype with new class features. 4.The case-based efficient knowledge perception incremental learning model building method of claim 1, wherein, The adaptive memory regularization module saves the classifier weights of the current t-1 task as an old memory unit in a memory pool after the end of each task During learning of a new task t, for the forward propagation process, the cross-entropy loss is used The incremental classifier is optimized, and the gradient p of the incremental classifier is used during back propagation in Momentum update is performed on each old memory unit, and a dynamic implicit memory loss L dhm (ξ) Regularize the classifier output with the aid of new knowledge.

5. The case-based high-efficiency knowledge perception incremental learning model establishing method according to claim 4, characterized in that, The dynamic implicit memory loss L dhm The calculation method of (ξ) is specifically: where ζ denotes a scaling factor, h is the batch size for the current task t phase, φ for the old memory cell selected from the memory pool, t-i f represents the parameters of the old memory cell, new denotes the sample of the new task passing through the feature extractor the resulting features, P(v p ) = (p(1,v p |f new ), p(2,v p |f new ),..., p(η t-i ,ν p |f new ) T denotes the output of the fine-tuned incremental classifier module, and the elements in p(i,ν p |f new ) are calculated as follows: wherein is the fine-tuning increment classifier for the t-th task, η t-i is the corresponding classifier dimension of. 6.The case-based efficient knowledge perception incremental learning model building method of claim 1, wherein, The prior guided incremental classifier module includes a fine-tuned incremental classifier and a self-supervised augmented classifier; wherein, in the t-th task stage, the fine-tuning incremental classifier inherits the weights from the previous stage classifier and copy it twice to obtain two classifiers equivalent to the old task class dimension For two classifiers, extend two fine-tuning components to adapt to new classes, respectively defined as and Add a projection function Proj(·) to the classifier , project the fine-tuning result to the new task dimension, and finally combine the outputs of the two classifiers to obtain The calculation formula is as follows: In the formula, Concat(·,·) represents the aggregation of two input features along the channel dimension, f is the sample through the feature extractor The obtained features, LoRA(·) is a fine-tuning component composed of two linear layers; The self-supervised augmented classifier is used to obtain output probability by putting the augmented image into it, and to enhance the generalization ability of the model by self-supervised contrast learning with the output probability of the fine-tuned incremental classifier. 7.The case-based efficient knowledge perception incremental learning model building method of claim 3, wherein, Step S4 comprises: Step S4-1: setting training parameters, if the current task is the base task, randomly initializing the efficient knowledge-aware incremental learning model based on no examples, if it is an incremental task, directly inheriting the network parameters obtained by training the last task to initialize; Step S4-2: training the t-th task dataset D t the i-th training image D t,i rotating 4 angles (0°, 90°, 180°, 270°) to obtain the rotated augmented image and calculating the cross-entropy loss and the self-supervised loss and optimizing the model; if the current task belongs to the incremental task, continue to execute step S4-3, otherwise, go to step S4-5; Step S4-3: Calculate the knowledge distillation loss between the network trained for the current task and the frozen network of the previous task ​​ Step S4-4: Obtain the old class prototype feature ξ by sampling from the distribution The old class prototype feature ξ is obtained by sampling from the distribution, and the dynamic implicit loss L is calculated through adaptive memory regularization dhm (ξ) Step S4-5: If the current task belongs to the base task, the cross-entropy loss calculated in step S4-2 is utilized and the self-supervised loss The network parameters are optimized, and the loss is calculated as follows: If the current task belongs to the incremental task, combine the cross-entropy loss Self-supervised loss Knowledge distillation loss Class separation loss And the dynamic implicit loss L dhm (ξ), optimize the network parameters, and the calculation method of the loss is as follows: In the formula, L is the total loss of the model, λ1 and λ2 are super parameters for controlling the importance of the loss respectively; Step S4-6: After the current stage, the prototype sampling method is used to update the prototype library for each class The prototype library is updated by the confidence prototype sampling method, and the current stage incremental classifier weight update memory pool is saved Wherein t represents the current task number, and h represents the number of old memory units in the memory pool. 8.The case-based efficient knowledge perception incremental learning model building method of claim 7, wherein, The self-supervised loss The calculation method of the self-supervised loss is specifically: where KL(·) is the KL-divergence loss, P denotes the probability distribution output by the classifier, f is the current stage D t where f aug is a self-supervised augmented classifier parameterized by κ.​​​​ 9.The case-based efficient knowledge perception incremental learning model building method of claim 7, wherein, The knowledge distillation loss The calculation method is specifically: where f old is the current stage output feature of the sample in the current stage through the old network feature extractor , ||·||2 is the L2 distance loss, and f is the feature of the sample in the current stage through the feature extractor the feature of the sample in the current stage through the feature extractor​ The knowledge distillation loss The calculation method is specifically: In the formula, d(·) is the Euclidean distance between two samples, a is an anchor sample, p is a positive sample, i.e. a sample with the same label as the anchor sample, n is a negative sample, i.e. a sample with a different label from the anchor sample, and m is a boundary parameter for forcing the expected separation between positive and negative samples.

10. A no-fewson-based efficient knowledge-aware incremental learning model apparatus, comprising a processor and a computer program stored on a memory and executable on the processor, characterized in that: The processor implements the method of any one of claims 1-9 when executing the program.

Images