A human-in-the-loop continuous learning method for strip surface defect classification

Abstract

The present application relates to a kind of human in loop strip surface defect classification continuous learning method, using the new class discovery strategy based on feature comparison makes model in line self-discovery new class data, and using human in loop new class identification strategy for new class data is labeled;When continuously learning to new class flow data, first using the flow data processing method based on sampling and dynamic buffering to process data, then using the continuous learning algorithm PDGAS based on prototype growth and population agent to learn new class data, expand the ability of model;The present application solves the problem that model cannot self-discover and label new class in industrial application, so that model is difficult to continuously learn, traditional industrial model cannot continuously grow, cannot cope with the problem of actual industrial flow data scene.

Description

Technical Field

[0001] This invention relates to the field of steel surface defect identification and classification technology, and in particular to a human-in-the-loop continuous learning method for classifying steel surface defects. Background Technology

[0002] Surface defects are common quality problems in strip steel production, often indicating a problem at some stage of the production line. Identifying the type of defect can, to some extent, reflect which stage of the production line the problem occurs in, facilitating timely repairs and reducing losses for the company. However, traditional defect classification methods often require extensive manual adjustments and optimizations, or the construction of deep learning models. But these deep learning models are usually static after training, with fixed model weights that cannot be adjusted further. Moreover, the types of defects the model can classify are fixed from the beginning, making it impossible to learn new defect types that appear on the production line except through retraining. In real-world engineering scenarios like strip steel surface defect classification, the continuously recorded images of the strip steel surface on the production line constitute a data stream reflecting the production line's status. Changes in materials, processing techniques, and other factors may lead to changes in defect characteristics, but static models cannot identify and classify newly generated defect types, making it difficult to cope with the diversity of strip steel surface defects and the constantly changing industrial environment. Therefore, the academic community has proposed continuous learning methods.

[0003] Current research on continuous learning methods can be mainly subdivided into replay-based, regularized, and parameter isolation methods. Replay-based continuous learning uses two modules: a learning module and a memory module. The learning module is mainly responsible for completing the learning task, while the memory module is responsible for replaying the required old data to the learning module. By storing old data and replaying it when the model learns new data, the model's memory of old knowledge is awakened, mitigating catastrophic forgetting and better completing the learning of new knowledge and updating the model's capabilities. Currently, it is the most effective method to combat catastrophic forgetting. Depending on whether a subset of samples is directly stored, it can be further divided into direct replay and generative replay. The direct replay method uses the previous samples stored in the memory buffer to fine-tune the network together with the current samples. The generative replay method generates old task data for replay-based continuous learning, but the replay method requires additional storage of old task data or distributions. As the number of learning tasks increases, the storage or computational costs caused by this type of method also increase continuously.

[0004] Regularization methods do not require storing data from old tasks, avoiding privacy issues and storage costs. Instead, they strengthen constraints on the loss function of the new task when updating model weights, thereby protecting and consolidating the learned knowledge from being forgotten. This maintains the model's memory of old knowledge without storing old data. Based on the object to which the constraints are applied, regularization can be divided into parametric regularization and distribution regularization. The main idea of ​​parametric regularization is to quantify the importance of model weights, select model weights highly correlated with old tasks, and train new tasks while minimizing the influence of these weights, thus suppressing catastrophic forgetting. Distribution regularization, on the other hand, is based on knowledge distillation loss, making the predictions of the old and new models similar on new tasks. However, this method cannot effectively constrain the model's optimization process on new tasks; excessive constraints on model parameters lead to poor learning performance on new tasks, resulting in poor model performance on new tasks.

[0005] Continuous learning methods based on parameter isolation address the catastrophic forgetting problem by assigning specific model parameters to each task. These methods are divided into dynamic architecture and static architecture methods. Dynamic architecture methods achieve continuous learning by freezing or growing the network, while static architecture methods only activate relevant parameters for each task without modifying the architecture. However, these methods explicitly associate neurons with different tasks, so task boundaries must be considered during model training and inference, which limits their application scenarios. Existing continuous learning methods all have certain limitations and are mostly designed for static offline scenarios, lacking practical value.

[0006] In addition, most current continuous learning methods do not consider real-world application scenarios. Streaming data on the surface of strip steel on industrial production lines has characteristics such as continuous data flow, data that has flowed through will not reappear, new types of data will appear, and the number of new and old types of data will be unbalanced. The data flow also exhibits a non-independent and identically distributed pattern, making it difficult for models to learn continuously from it. Summary of the Invention

[0007] In view of this, in order to address the problems of poor practicality and difficulty in handling streaming, class imbalance, and non-independent and identically distributed data in real-world industrial production scenarios using traditional continuous learning methods, and to introduce the strip steel surface defect classification scenario to solve the problem that traditional defect classification models cannot identify newly emerging defect types, this invention provides a human-in-the-loop continuous learning method for strip steel surface defect classification, comprising the following steps:

[0008] S1. Acquire flow cytometry data of surface defects in strip steel on industrial production lines;

[0009] S2. The streaming data of surface defects of the strip steel is processed using a sampling and dynamic buffering-based streaming data processing method, including the reservoir sampling method, the memory buffer dynamic allocation method, and the random retrieval strategy.

[0010] S201. Employ a new class discovery strategy based on feature comparison to determine the newly acquired streaming data x. j Does j≥1 belong to a category that has already appeared?

[0011] By calculating the input streaming data x j The cosine similarity Ψ(x) between the projection vector in the feature space and the category prototypes stored in the strip surface defect classification model. j ),

[0012] If Ψ(x) j If )>γ, then the strip surface defect classification model determines the lost input data x. j If it belongs to an existing category, proceed to step S202, where γ represents a preset threshold.

[0013] If Ψ(x) j If ) < γ, then the strip surface defect classification model determines the lost input data x. j If it belongs to a category that has not appeared, proceed to step S203 and report it to the data identification platform;

[0014] S202. Create new category buffers and adjust the size of each category buffer using the dynamic allocation method of memory buffers;

[0015] S203. A novel identification strategy for human-in-the-loop systems is designed, employing a reservoir sampling method to determine whether newly acquired streaming data x should be included. j Add it to the corresponding category memory buffer;

[0016] S204. Using a random retrieval strategy, retrieve an approximately independent and identically distributed replay training batch from the balanced memory buffer and input it into the strip steel surface defect classification model.

[0017] S3. Optimize the strip steel surface defect classification model using the continuous learning algorithm PGAS;

[0018] S301. Based on the characteristic information of streaming data batches, dynamically calculate, update and store category prototypes using prototype storage and growth strategies;

[0019] S302. Establish attraction and repulsion sets for streaming data used in continuous learning through a population proxy strategy;

[0020] S303. Promote the optimization of the feature space of the classifier network through the group proxy loss;

[0021] S304. Update the strip steel surface defect classification model.

[0022] Preferably, the cosine similarity ψ(x) j The expression is:

[0023]

[0024] Where, x j p represents the input streaming data. i f represents the category prototype. θ (x j ) represents the projection vector of the input streaming data onto the feature space.

[0025] Preferably, the water storage tank sampling method involves constructing a water storage tank of size k, where the surface defect flow cytometry data x of the strip steel is... j If position j ≤ k in the data stream, then j of the strip surface defect streaming data directly enter the storage tank; if the strip surface defect streaming data x j If the position j>k in the data stream, then a random number i∈[1,j] is generated. When i≤k, then the streaming data x of the strip surface defects is... j The data is directly fed into the reservoir and replaces the existing i-th data. If i ≤ k, the data x is discarded. j .

[0026] Preferably, the dynamic allocation method for the memory buffer includes: if newly acquired data x j If the data belongs to a category that has already appeared, the reservoir sampling method is used to determine whether to store it in the memory buffer of the corresponding category; if the newly acquired data x j If a category has not appeared before, the memory buffer size allocated to each category is recalculated, the memory buffer size of all existing categories is adjusted, samples are deleted based on the current number of cached samples, and a new category memory buffer is created. Newly acquired data (x) is then sampled using a reservoir sampling method. j Move to the new category memory buffer;

[0027] The expression for the memory buffer size allocated to each category is:

[0028]

[0029] Where m represents the size of the memory buffer allocated to each category, K represents the total stream data of surface defects in the strip steel currently acquired, and t represents the number of categories that have appeared so far.

[0030] Preferably, the novel class identification strategy based on human-in-the-loop is based on the expertise and experience of human experts. It uses a visual interface to label data that does not belong to an existing class with the correct class label, and then transmits the labeled data to the strip surface defect classification model for continuous learning through a data transmission interface.

[0031] Preferably, the prototype storage and growth strategy stores newly acquired streaming data in batches to the closest category prototype p during the dynamically changing streaming process. c The expression is:

[0032]

[0033] Where μ∈[0,1] represents the parameter controlling the update magnitude. The mean feature value of class c samples in the feature space in the current training streaming data batch is expressed as:

[0034]

[0035] Among them, B c ={(x i ,y i =c)∈B} represents all sample streaming data in category c contained in the current training streaming data batch B, |B c | indicates the number of sample streaming data in category c within the current training streaming data batch B. This represents the sample streaming data contained in category c. Represents sample streaming data The corresponding feature vector.

[0036] Preferably, in the training streaming data batch B, select the sample streaming data. Other sample streaming data in the same category Sample streaming data The projection mean on the feature space serves as the population proxy. The ethnic group agent and category prototype p c Form an attraction set The expression is:

[0037]

[0038] Sample streaming data Projection on feature space And the subset of category c is used as the exclusion set for other categories in the training streaming data batch B. The expression is:

[0039]

[0040] in, This represents sample streaming data from other categories.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] 1. The method provided by this invention processes the streaming data of strip steel surface defects on the industrial production line through water tank sampling, dynamic allocation of memory buffer and random retrieval strategy, which enhances the quality of model learning category feature space and improves the continuous learning effect of strip steel surface defect classification model.

[0043] 2. This invention processes data using a streaming data processing method based on sampling and dynamic buffering, and completes autonomous discovery and labeling of new category data and learning using the PGAS continuous learning algorithm based on prototype growth and population proxy, thereby expanding the model's capabilities and realizing a semi-self-help lifelong growth function.

[0044] 3. The method provided by this invention solves the problem that traditional strip steel surface defect classification models cannot identify new defect types, thereby improving industrial production efficiency and product quality. Attached Figure Description

[0045] Figure 1 This is a flowchart of a human-in-the-loop continuous learning method for classifying surface defects in strip steel according to the present invention.

[0046] Figure 2 This is a flowchart of the data processing of the present invention;

[0047] Figure 3 This is a flowchart of the PGAS continuous learning algorithm of this invention;

[0048] Figure 4 This is a schematic diagram of the water storage tank sampling method of the present invention;

[0049] Figure 5 This is a schematic diagram of the dynamic allocation strategy for the memory buffer of the present invention;

[0050] Figure 6 This is a schematic diagram of the random retrieval strategy of the present invention;

[0051] Figure 7 This is a schematic diagram of the novel identification strategy for in-circuit testing invented by the inventors. Detailed Implementation

[0052] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0053] Example 1

[0054] Most current continuous learning methods do not consider practical application scenarios. Streaming data from industrial production lines exhibits characteristics such as continuous data flow (data that has passed through will not reappear), the emergence of new data types with an imbalance in the number of new and old categories, and a non-independent, identically distributed data stream. These characteristics make it difficult for models to continuously learn from these data. To address the problem that traditional defect classification models cannot identify newly emerging defect types, this invention provides a human-in-the-loop continuous learning method for classifying defects on the surface of strip steel, comprising the following steps:

[0055] S1. Acquire flow cytometry data of surface defects in strip steel on industrial production lines;

[0056] S2. The streaming data of surface defects of the strip steel is processed using a sampling and dynamic buffering-based streaming data processing method, including the reservoir sampling method, the memory buffer dynamic allocation method, and the random retrieval strategy.

[0057] S201. Employ a new class discovery strategy based on feature comparison to determine the newly acquired streaming data x. j Whether j≥1 belongs to a category that has already appeared. The new class discovery strategy enables the model to autonomously discover new classes on the production line and report the image data of these new classes.

[0058] By calculating the input streaming data x j The cosine similarity Ψ(x) between the projection vector in the feature space and the category prototypes stored in the strip surface defect classification model. j The expression is:

[0059]

[0060] Where, x j p represents the input streaming data. i f represents the category prototype. θ (x j () represents the projection vector of the input streaming data onto the feature space;

[0061] If Ψ(x) j If )>γ, then the strip surface defect classification model determines the lost input data x. j If it belongs to an existing category, proceed to step S202, where γ represents a preset threshold.

[0062] If Ψ(x) j If ) < γ, then the strip surface defect classification model determines the lost input data x. j If it belongs to a category that has not appeared, proceed to step S203 and report it to the data identification platform;

[0063] like Figure 4As shown, the reservoir sampling method constructs a reservoir of size k, and if the surface defect flow cytometry data x of the strip steel is... j If position j ≤ k in the data stream, then j of the strip surface defect streaming data directly enter the storage tank; if the strip surface defect streaming data x j If the position j>k in the data stream, then a random number i∈[1,j] is generated. When i≤k, then the streaming data x of the strip surface defects is... j The data is directly fed into the reservoir and replaces the existing i-th data. If i ≤ k, the data x is discarded. j The reservoir sampling method ensures that all data in each class have an equal probability of being selected during the data flow process, thus making the samples in the memory buffer more representative.

[0064] S202. Create new category buffers and adjust the size of each category buffer using the dynamic allocation method of memory buffers;

[0065] like Figure 5 As shown, the dynamic allocation method for the memory buffer includes: if newly acquired data x j If the data belongs to a category that has already appeared, the reservoir sampling method is used to determine whether to store it in the memory buffer of the corresponding category; if the newly acquired data x j If a category has not appeared before, the memory buffer size allocated to each category is recalculated, the memory buffer size of all existing categories is adjusted, samples are deleted based on the current number of cached samples, and a new category memory buffer is created. Newly acquired data (x) is then sampled using a reservoir sampling method. j The memory buffer is dynamically allocated to each observed category using a dynamic memory buffer allocation strategy. This ensures that an equal amount of data is cached for each category, thus creating a balanced memory buffer.

[0066] The expression for the memory buffer size allocated to each category is:

[0067]

[0068] Where m represents the size of the memory buffer allocated to each category, K represents the total stream data of surface defects in the strip steel currently acquired, and t represents the number of categories that have appeared so far;

[0069] S203. A novel identification strategy for human-in-the-loop systems is designed, employing a reservoir sampling method to determine whether newly acquired streaming data x should be included. j Add it to the corresponding category memory buffer;

[0070] like Figure 7As shown, the novel human-in-the-loop classification strategy is based on the expertise and experience of human experts. It uses a visual interface to label data that does not belong to the existing categories with the correct category labels, and then transmits the labeled data to the strip surface defect classification model for continuous learning through a data transmission interface.

[0071] S204. Using a random retrieval strategy, retrieve an approximately independent and identically distributed replay training batch from the balanced memory buffer and input it into the strip surface defect classification model, such as... Figure 6 As shown, the random retrieval strategy allows the model to learn from a more diverse range of samples, thereby improving the model's learning performance.

[0072] S3. Optimize the strip steel surface defect classification model using the continuous learning algorithm PGAS;

[0073] S301. Based on the characteristic information of streaming data batches, dynamically calculate, update and store category prototypes using prototype storage and growth strategies;

[0074] The prototype storage and growth strategy stores newly acquired streaming data in batches to the closest category prototype p during a dynamically changing streaming process. c The expression is:

[0075]

[0076] Where μ∈[0,1] represents the parameter controlling the update magnitude. The mean feature value of class c samples in the feature space in the current training streaming data batch is expressed as:

[0077]

[0078] Among them, B c ={(x i ,y i =c)∈B} represents all sample streaming data in category c contained in the current training streaming data batch B, |B c | indicates the number of sample streaming data in category c within the current training streaming data batch B. This represents the sample streaming data contained in category c. Represents sample streaming data The corresponding feature vector;

[0079] Prototype storage and growth strategies address issues that data processing methods cannot resolve, such as the changing distribution of similar streaming data over time.

[0080] S302. Establish attraction and repulsion sets for streaming data used in continuous learning through a population proxy strategy;

[0081] Select the sample streaming data in training streaming data batch B. Other sample streaming data in the same category Sample streaming data The projection mean on the feature space serves as a population proxy. Ethnic Agents and category prototype p c Form an attraction set The expression is:

[0082]

[0083] Sample streaming data Projection on feature space And the subset of category c is used as the exclusion set for other categories in the training streaming data batch B. The expression is:

[0084]

[0085] in, This represents streaming data representing samples from other categories;

[0086] The group proxy strategy addresses the problem that the streaming data of steel surface defects has large intra-class differences and small inter-class differences, which makes it difficult for the model to form a feature space with a clear class distribution, thus reducing the classification performance of the model. Based on the attraction and repulsion sets in the group proxy, the inter-class differences are effectively expanded in the feature space while the intra-class differences are reduced, thereby improving the classification performance of the model.

[0087] S303. Promote the optimization of the feature space of the classifier network through the group proxy loss;

[0088] S304. Update the strip steel surface defect classification model;

[0089] In this embodiment, the PDGAS continuous learning algorithm based on prototype growth and population proxy is used to learn new categories of data, expand the model's capabilities, and solve problems such as the inability of models to autonomously discover and label new categories in industrial applications, the inability of traditional industrial models to continuously grow, and the inability of existing continuous learning algorithms to cope with actual industrial streaming data scenarios. Finally, a semi-autonomous lifelong growth strip steel surface defect classification system solution is obtained.

[0090] Example 2: Online Continuous Learning of Strip Steel Surface Defects

[0091] Based on image data of surface defects in strip steel generated on actual industrial production lines, a dataset of surface defects in strip steel was constructed, and an experiment was designed to closely resemble the actual industrial production scenario of strip steel.

[0092] The dataset is based on images of surface defects in strip steel generated on real industrial production lines. It contains 7,700 data samples across 10 categories, each sample being a 780×128 grayscale image. The ratio of training data to test data is 4:1. There is a significant difference in the number of samples among different defect categories in the dataset, with the category with the most samples having more than twice the number of samples as the category with the fewest samples. This imbalance in the number of samples among categories is a characteristic of data from real-world scenarios and presents a significant challenge to continuous learning algorithms.

[0093] Following the standard experimental design in the field of continuous learning, the strip steel surface defect dataset was randomly divided into 5 learning tasks according to categories. To simulate a real industrial scenario, the dataset was processed into streaming data, and an algorithm was used for online continuous learning. Representative methods in the field of continuous learning were selected for comparative experiments. The experimental results are shown in Table 1.

[0094] Table 1 shows the experimental results of each algorithm on the strip surface defect dataset.

[0095]

[0096] Table 1 shows that the algorithm proposed in this invention achieves the best final average accuracy on a real strip surface defect dataset, while also taking less time.

[0097] Example 3: Discovery of New Classification Scenarios for Strip Steel Surface Defects

[0098] To verify the effectiveness and performance of the proposed novel class discovery strategy based on feature comparison in the strip steel surface defect classification scenario, an experimental scenario for novel class discovery was constructed. The proposed continuous learning algorithm PGAS was used to learn data from half of the categories on the strip steel surface defect dataset. The algorithm was then tested on a test set containing all categories. The accuracy of the novel class discovery strategy was evaluated by dividing the number of correctly classified samples (including those correctly classified as known categories and those correctly classified as new categories) by the total number of samples in the test set. Considering that the threshold γ is a parameter that directly affects the recognition accuracy in the novel class discovery strategy, it is necessary to test the performance of the strategy under different threshold γ values. The experimental results are shown in Table 2.

[0099] Table 2. Experimental results of the novel class discovery strategy in the strip surface defect classification scenario.

[0100]

[0101]

[0102] Analysis of Table 2 reveals that the new class discovery strategy performs well when the threshold value is reasonable. The threshold γ is a crucial factor affecting strategy performance: if the threshold γ is too high, even data within previously learned categories may be incorrectly classified as new categories; conversely, if the threshold is set too low, data belonging to new categories may be incorrectly classified into known categories, thus degrading the strategy's effectiveness. Therefore, selecting an appropriate threshold γ is key to fully leveraging the strategy's capabilities.

[0103] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A continuous learning method for classifying surface defects in strip steel using a human-in-the-loop system, characterized in that, Includes the following steps: S1. Acquire flow cytometry data of surface defects in strip steel on industrial production lines; S2. The streaming data of surface defects of the strip steel is processed using a sampling and dynamic buffering-based streaming data processing method, including the reservoir sampling method, the memory buffer dynamic allocation method, and the random retrieval strategy. S201. Employ a new class discovery strategy based on feature comparison to determine newly acquired streaming data. Does it belong to a category that has already appeared? By calculating the input streaming data The cosine similarity between the projection vector in the feature space and the category prototypes stored in the strip surface defect classification model. , like Then the strip surface defect classification model determines the lost input data. If it belongs to an existing category, proceed to step S202, where... Indicates the preset threshold; like Then the strip surface defect classification model determines the lost input data. If it belongs to a category that has not appeared, proceed to step S203 and report it to the data identification platform; S202. Create new category buffers and adjust the size of each category buffer using the dynamic allocation method of memory buffers; S203, a novel identification strategy for human-in-the-loop systems, employing a reservoir sampling method to determine whether newly acquired streaming data should be included. Add it to the corresponding category memory buffer; The novel human-in-the-loop identification strategy is based on the expertise and experience of human experts. It uses a visual interface to label data that does not belong to an existing category with the correct category label, and then transmits the labeled data to the strip surface defect classification model for continuous learning through a data transmission interface. S204. Using a random retrieval strategy, retrieve an approximately independent and identically distributed replay training batch from the balanced memory buffer and input it into the strip steel surface defect classification model. S3. Optimize the strip steel surface defect classification model using a continuous learning algorithm; S301. Based on the characteristic information of streaming data batches, dynamically calculate, update and store category prototypes using prototype storage and growth strategies; The prototype storage and growth strategy stores newly acquired streaming data in batches to the closest category prototype during the dynamically changing streaming process. The expression is: in, This parameter represents the parameter that controls the update magnitude. Indicates the category in the current training streaming data batch The feature mean of the sample in the feature space is expressed as: in, Indicates the current training streaming data batch Medium category All sample streaming data included. Indicates the current training streaming data batch Medium category The number of sample streaming data included. Indicates category The sample streaming data included Represents sample streaming data The corresponding feature vector; S302. Establish attraction and repulsion sets for streaming data used in continuous learning through a population proxy strategy; S303. Promote the optimization of the feature space of the classifier network through the group proxy loss; S304. Update the strip steel surface defect classification model.

2. The continuous learning method for classifying surface defects of strip steel in a human-in-the-loop manner according to claim 1, characterized in that, The cosine similarity The expression is: in, This represents the input streaming data. Represents the category prototype. This represents the projection vector of the input streaming data onto the feature space.

3. The continuous learning method for classifying surface defects of strip steel in a human-in-the-loop manner according to claim 1, characterized in that, The water storage tank sampling method involves constructing a tank of size... The water storage tank, if the surface defect data of the strip steel is flow cytometry Position in the data stream ,but The flow cytometry data of surface defects in the strip steel is directly fed into the water storage tank; if the flow cytometry data of surface defects in the strip steel is... Position in the data stream Then a random number is generated. When satisfied When, the strip surface defect streaming data Directly enter the reservoir and replace the original first Data, when not satisfied At that time, discard data .

4. The continuous learning method for classifying surface defects of strip steel in a human-in-the-loop manner according to claim 1, characterized in that, The dynamic allocation method for the memory buffer includes: if newly acquired data If the data belongs to a category that has already appeared, the reservoir sampling method is used to determine whether to store it in the memory buffer of the corresponding category; if the data is newly acquired... For categories that have not appeared before, the memory buffer size allocated to each category is recalculated, the memory buffer size of all existing categories is adjusted, samples are deleted based on the current number of cached samples, and new category memory buffers are created. Newly acquired data is then sampled using a reservoir sampling method. Move to the new category memory buffer; The expression for the memory buffer size allocated to each category is: in, This indicates the size of the memory buffer allocated to each category. , This indicates the number of categories that have appeared so far.

5. The continuous learning method for classifying surface defects of strip steel in a human-in-the-loop manner according to claim 1, characterized in that, In training streaming data batches Select with sample streaming data Other sample streaming data in the same category Streaming sample data , The projection mean on the feature space serves as a population proxy. The group agent and category prototype Form an attraction set The expression is: Sample streaming data Projection on feature space and its category as training streaming data batch Exclusion sets of other categories The expression is: in, This represents sample streaming data from other categories.

Images