Class increment classification method based on union search set graph modeling and fusion space

Abstract

The invention discloses a class increment classification method based on union search set graph modeling and fusion space, which comprises the following steps: configuring an independent adaptation module for each increment stage, extracting class characteristics of each stage by freezing a pre-training visual backbone network, and constructing a stage class prototype set. And calculating cross-stage similarity among global category prototypes, establishing a category confusion relationship, and identifying confusion connected branches by using a union-check set structure. And for each connected branch, fusing the features of the adaptive module in the corresponding increment stage and optimizing the fusion weight to form a fusion feature space. In the inference stage, the preliminary classification result is used for judging whether the candidate category belongs to a confusion connected branch or not, if yes, fine judgment is carried out in the fusion feature space to output a final result, and if not, the preliminary classification result is directly output. The method effectively solves the identification problem of cross-stage easily-confused categories, improves the classification precision, and maintains the low incremental training cost and the anti-forgetting capability of the model at the same time.

Description

Technical Field

[0001] This invention belongs to the fields of machine learning and computer vision technology, specifically relating to an incremental classification method for categories based on disjoint-set graph modeling and fusion space. Background Technology

[0002] Deep learning models have achieved outstanding performance in computer vision classification tasks, but traditional models typically rely on batch training of fixed categories, making it difficult to adapt to the need for continuously increasing categories in real-world scenarios. Class-Incremental Learning (CIL), as a key technology to solve this problem, allows models to progressively learn new category knowledge while retaining the ability to learn already learned categories, and has significant value in practical applications such as intelligent monitoring, autonomous driving, and medical diagnosis.

[0003] In recent years, continuous learning based on pre-trained models has shown significant potential for efficient adaptation across task sequences, becoming a key foundation for mitigating the catastrophic forgetting problem in incremental learning. To reduce the cost of incremental training, the strategy of using lightweight auxiliary modules, especially adapter fine-tuning, has gained widespread attention and application. This method freezes the parameters of the pre-trained backbone network and only fine-tunes a small number of newly added adapters to adapt to new tasks. This fully utilizes the feature extraction capabilities of the pre-trained model while effectively mitigating the risk of catastrophic forgetting caused by parameter updates.

[0004] While adapter fine-tuning performs exceptionally well in incremental learning, existing methods still have significant limitations. Adapters trained at different incremental stages typically focus only on learning category features for the current stage, failing to establish an effective mechanism for distinguishing similar categories across stages. Specifically, the prototype category features extracted by the corresponding adapter in a previous stage may have high similarity in the feature space to the prototypes of similar categories extracted by dedicated adapters in subsequent stages. This makes it difficult for the model to accurately distinguish these easily confused categories across stages during inference, severely impacting overall classification performance. This similarity information could actually be used to guide further discriminative modeling of features of similar categories across stages.

[0005] Furthermore, prior art CN121305230A discloses a few-shot incremental classification method based on global and local collaboration. Its core idea is to alleviate few-shot overfitting through global alignment of visual and textual prototypes, local alignment of color and shape attributes, and knowledge distillation-guided global and local textual cue learning. However, this method has limitations: it lacks modeling for cross-stage category confusion. While prior art employs global and local cue learning, it focuses on addressing overfitting in few-shot classification and does not specifically address category confusion in incremental learning. Especially between categories in multiple incremental stages, similar or easily confused categories may fail to be effectively distinguished. This can lead to misclassification during inference in practical applications, particularly when category similarity exists between incremental stages.

[0006] Therefore, in application scenarios where categories are continuously increasing, how to structurally model the similarity between categories across stages and make targeted distinctions between potentially confusing categories during the inference stage remains a technical problem that needs to be further solved. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a category incremental classification method based on disjoint-set graph modeling and fusion space.

[0008] The objective of this invention can be achieved through the following technical solutions: This invention provides a method for incremental category classification based on disjoint-set graph modeling and fusion space, comprising the following steps: Construct an incremental learning task environment, load a pre-trained visual backbone network and freeze its parameters, and configure an independent adaptation module for each incremental stage in the visual backbone network. According to the preset incremental stage order, the corresponding adaptation module of the incremental stage is trained sequentially using the newly added category data of the corresponding incremental stage. After the training is completed, its parameters are fixed, and the feature representation of each category of the incremental stage is extracted based on the trained visual backbone network and the corresponding adaptation module to construct the category prototype set of the incremental stage. After completing all incremental training phases, the class prototype sets of all incremental phases are aggregated to form a global class prototype set. Based on the global category prototype set, the similarity relationship between category prototypes from different incremental stages is calculated. When the preset confusion judgment condition is met, the confusion association between categories is established, and a cross-stage category confusion relationship set is constructed. Based on the set of category confusion relationships, a cross-stage category confusion graph is constructed with each category as a node and the confusion association as an edge. The confusion graph is then processed by a disjoint-set data structure to obtain several category confusion connected branches. For the confused connected branches, based on the incremental stage to which each category belongs within the branch, the corresponding adaptation module is selected to construct the fusion feature space, and the ability to distinguish between categories within the confused connected branches is enhanced by optimizing the fusion weights, thereby obtaining the fusion feature space model corresponding to each confused connected branch. During the inference phase, for the sample to be classified, the sample features are extracted through the adaptation modules of each incremental phase and matched with the global category prototype set to obtain preliminary classification results. If the preliminary classification results contain candidate categories belonging to the same confused connected component but from different incremental stages, the corresponding fusion feature space model is invoked to further distinguish the candidate categories and output the final classification result; otherwise, the preliminary classification result is directly output as the final classification result.

[0009] Furthermore, the construction of the incremental learning task environment specifically includes: The incremental learning task is defined as having a total number of incremental stages. The datasets corresponding to each incremental stage are as follows: Where the datasets in any two different incremental stages do not overlap in categories, satisfying ,in ; Dataset for each incremental stage This includes the set of training samples for the newly added categories in this incremental phase and their corresponding set of category labels, specifically represented as follows: in, Indicates the first The first incremental stage One input sample, wherein the input sample is image data; This represents the category label corresponding to the input sample; Indicates the first The total number of samples in each incremental phase; The dataset for the incremental stage is used to supervise the training of the adaptation module for the corresponding incremental stage, so that the adaptation module learns the feature representation of the newly added category in the stage under the condition of freezing the visual backbone network parameters, and is used to calculate the category prototype of each category in the stage after training is completed.

[0010] Furthermore, the loading of the pre-trained visual backbone network and freezing of its parameters, and the configuration of an independent adaptation module for each incremental stage in the visual backbone network, specifically includes: Load the pre-trained visual backbone network and freeze all its network parameters so that the weights are not updated during each incremental training phase. An adapter module is inserted into each feature transformation layer of the visual backbone network. , Each incremental stage corresponds to an independent adaptation module, which adopts a bottleneck structure and includes a dimensionality reduction mapping matrix. Nonlinear activation functions and up-dimensional mapping matrices ; The adaptation module is inserted into the feedforward network bypass of each layer of the visual backbone network in the form of residual connections, and the output of the layer feedforward network is corrected as follows: in, This represents the input feature representation of the corresponding layer in the visual backbone network. The output of the feedforward network after the integration and adaptation module; This represents the feedforward neural network module of this layer in the visual backbone network; Represents the linear rectification activation function; Initialize the corresponding adaptation module parameters for each incremental stage. The parameters of the adaptation modules in each incremental stage are independent and not shared, and are used to learn the feature representation of the corresponding stage category at different stages.

[0011] Furthermore, the step of training the adaptation module corresponding to the incremental stage using the newly added category data of the corresponding incremental stage in a predetermined incremental stage order specifically includes: For the In the first incremental stage, the parameters of the visual backbone network are fixed, and only the first... The corresponding adaptation module for each incremental stage parameters Set to trainable state; The dataset of this incremental phase The input consists of a visual backbone network and an adaptation module. The feature extraction model is composed of these components to obtain the feature representation of the samples. A classification prediction function is constructed based on the feature representation, and the prediction output is obtained through a linear classifier. A supervised loss function is constructed based on the true class labels and predicted outputs in the dataset, and the loss function adopts cross-entropy loss; The loss function is minimized using the gradient descent optimization method, updating only the parameters of the adaptation module and the classifier parameters of the corresponding stage, until the preset convergence condition is met or the maximum number of training rounds is reached.

[0012] Furthermore, the step of extracting feature representations of each category in the incremental stage based on the trained visual backbone network and corresponding adaptation modules, and constructing a category prototype set for the incremental stage, specifically includes: In the After each incremental training phase is completed, the adaptation module is fixed. After setting the parameters, remove or freeze the classifier for the corresponding stage, retaining only the frozen visual backbone network and the first stage. Each incremental stage adaptation module serves as a feature extraction model; The dataset corresponding to the incremental stage The input samples are re-inputted into the feature extraction model to obtain the feature representation of each sample: in, Indicates the first In the first incremental stage Feature representation of each sample; This indicates that the input sample is frozen and connected to the visual backbone network. The feature mapping function after processing by the incremental stage adaptation module; Indicates the first The first incremental stage One input sample; Indicates the first The corresponding adaptation modules for each incremental stage; For the Any category in the incremental phase Select the category labels that meet the criteria. The feature representations of all samples are then aggregated by mean to obtain the category prototype for that category: in, Indicates the first Categories in each incremental stage The category prototype; Indicates the first Categories in each incremental stage The number of samples; Indicates the first The total number of samples in each incremental phase; This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise; Indicates the first In the first incremental stage The category labels of each sample; Indicates the first A set of categories for each incremental stage; The first The category prototypes corresponding to all categories in each incremental stage constitute the category prototype set for that incremental stage: in, Indicates the first A set of category prototypes for each incremental stage.

[0013] Furthermore, based on the global category prototype set, the similarity relationship between category prototypes from different incremental stages is calculated. When a preset confusion judgment condition is met, a confusion association between categories is established, and a cross-stage category confusion relationship set is constructed. Specifically, this includes: For any two categories in the global category prototype set, arbitrarily select two category prototypes from different incremental phases from the global category prototype set. and When satisfied Then, calculate the prototype similarity between the two: in, Indicates category With category Prototype similarity; Indicates the first Categories in each incremental stage The category prototype; Indicates the first Categories in each incremental stage The category prototype; The L2 norm of a vector; Set a preset confusion threshold When satisfied When determining the category With category Confusional relationships exist in the feature space; confusional edges are established between categories, denoted as... Otherwise, record as ;in, Indicates category With category The confusion between the indicator variables, when A '-' indicates that a confusing association is established; otherwise, a confusing association is not established. All satisfied and The category pairs constitute a set of cross-stage category confusion relationships: in, This represents a set of cross-stage category confusion relationships; This indicates a confusion association pair that is determined to have a confusion relationship between different incremental stages.

[0014] Furthermore, based on the set of category confusion relationships, a cross-stage category confusion graph is constructed with each category as a node and the confusion association as an edge. A disjoint-set data structure is then used to perform connectivity merging on the confusion graph, resulting in several category confusion connectivity branches, specifically including: Using categories from the global category prototype set as nodes in the graph, a set of category confusion relationships is created. Each pair of confusing associations As edges of the graph; obtain a cross-stage category confusion graph. , is represented as: in, The set of nodes in the graph, including category nodes for all incremental stages. ; The set of edges of the graph, including the set of category confusion relations. E All obfuscated association pairs; Initialize the disjoint-set data structure, setting each category as an independent set. The initial state of the disjoint-set is that each category node is a separate connected component. For category nodes The corresponding initial set is: in, Represents category nodes The parent node, Represents category nodes The depth of the tree; Traversing the confusion graph Each edge in For each pair of obfuscated associations, perform the following operations: For each pair of category nodes and Use the Find operation of the disjoint-set data structure to search for each element. and The root node. If Then, a Union operation is performed to merge the sets containing nodes of the two categories, and the corresponding updates are made. and information: After processing all edges, a disjoint set containing several connected components is obtained, where each connected component represents a set of categories connected in the graph through confusion relations; Each independent set in the disjoint-set data structure is divided into several obfuscated connected branches. Each obfuscated connected branch consists of multiple category nodes, which are interconnected through obfuscation relationships. The confused connected branches are filtered to select those containing at least two categories of nodes, which are then used as the final category-confused connected branches.

[0015] Furthermore, for the confused connected component, based on the incremental stage to which each category belongs within the component belongs, a corresponding adaptation module is selected to construct a fusion feature space. The ability to distinguish between categories within the confused connected component is enhanced by optimizing the fusion weights, thereby obtaining a fusion feature space model corresponding to each confused connected component. Specifically, this includes: For each filtered confusing connected component ,in, Let be the set of nodes in the graph; Based on each category node The incremental stage to which it belongs Select the corresponding pre-trained and fixed-parameter adaptation module. As a module for feature extraction in this category; Based on the confusion of all categories of nodes within the connected component Corresponding to the incremental phase, an adaptation module is issued. Define the fusion feature space of confused connected components. A weighted combination of features for the selected adaptation modules: in, Indicates confusing connected components The fusion feature space; Indicates confusing connected components The number of nodes in each category; Indicates the corresponding adaptation module The learnable fusion weight coefficients are used to adjust the contribution of features in each incremental stage to the fusion space; Indicates the first Adaptation modules for each incremental stage; Initialize fusion weight coefficients And satisfy the constraints: This will obfuscate the category prototypes of various categories within the connected component. Mapping to the fused feature space The fused feature prototype is obtained. ; In the fusion feature space In the middle, calculate the feature prototypes of all categories within the confusing connected component. Pairwise Euclidean distance between them: in, and , Indicates Euclidean distance; Based on the feature prototype The pairwise Euclidean distance between them defines the optimization objective of confusing connected components. : To maximize the optimization objective To achieve the objective, the gradient descent method is used to optimize the fusion weight coefficients. Iterate and update until convergence or the preset maximum number of iterations is reached, and save the optimized fusion weight coefficients and the corresponding fusion feature space description information to form confused connected components. Fusion feature space model .

[0016] Furthermore, during the inference phase, for the sample to be classified, sample features are extracted through the adaptation modules of each incremental phase and matched with the global category prototype set to obtain preliminary classification results, specifically including: Samples to be classified x The frozen visual backbone network is input and forward-propagated through the adaptation modules of each incremental stage to obtain the sample features of each stage: in, Indicates the sample to be classified After freezing the visual backbone network and the first Incremental Stage Adaptation Module Features of the processed samples; Indicates the total number of incremental stages; For each incremental stage Calculate sample features Cosine similarity with each category prototype in the incremental stage category prototype set ; The similarity calculation results from all stages are aggregated to generate a global similarity ranking list, and the top similarity scores are selected. Each category is used as a candidate category set. : Each candidate category contains a category identifier. Belonging to the incremental stage and corresponding similarity scores ; Indicates sample With incremental stage category Cosine similarity; Candidate category set This serves as a preliminary classification result.

[0017] Furthermore, when the preliminary classification result contains candidate categories belonging to the same confused connected component but from different incremental stages, the corresponding fusion feature space model is invoked to further discriminate the candidate categories, and the final classification result is output; if none exist, the preliminary classification result is directly output as the final classification result, specifically including: Examine the attribution relationship between the top-1 candidate category and other candidate categories in the preliminary classification results; if there is at least one candidate category... The second level of fine-tuning is triggered when the following two conditions are met: (a) It belongs to the same confusing connected component as the top-1 candidate category; (b) The top-1 candidate category comes from a different incremental phase; When the second level of fine-tuning is triggered, the obfuscated connectivity branch is invoked. Fusion feature space model The samples to be classified The features are mapped to the fused feature space to obtain the samples to be classified. Enhanced feature representation in the fused feature space; In the fusion feature space In the middle, calculate the samples to be classified Enhanced feature representation and obfuscated connected components Feature prototypes of all categories Cosine similarity; The category with the highest cosine similarity is used as the sample to be classified. The final predicted category; If the second-level fine-tuning condition is not triggered, the top-1 candidate category of the preliminary classification result will be directly output as the final predicted category.

[0018] Compared with the prior art, the present invention has the following advantages: (1) To address the problem of the lack of a unified category relationship modeling mechanism between different incremental stages in existing incremental learning methods, especially the technical defect that cross-stage similar categories are not structurally identified during multi-stage continuous learning, resulting in unclear discrimination boundaries in the inference stage, this invention constructs a graph structure model based on category prototype similarity, abstracts categories from different incremental stages into graph nodes, and establishes confusion connections between categories based on cross-stage prototype similarity. Furthermore, it utilizes a disjoint-set data structure to perform connectivity analysis on the category relationship graph, automatically identifying connected branches of cross-stage confusion categories, thus achieving structured grouping and global modeling of potential cross-stage confusion categories. This technical feature enables a unified characterization of previously scattered similar category relationships, avoiding the instability caused by relying solely on local similarity thresholds, improving the systematicity and completeness of cross-stage category relationship analysis, providing a clear structural foundation for subsequent targeted differentiation processing, and thereby enhancing the overall discrimination capability of the model in multi-stage category continuous expansion scenarios.

[0019] (2) To address the problem of independent adapters and fragmented feature spaces in existing adapter-based incremental learning methods, which leads to a lack of comparability and insufficient distinguishability of similar categories across different feature spaces, this invention selects the corresponding adaptation modules for each confused connected branch identified by disjoint-set data structure. While keeping the parameters of the adaptation modules frozen, a weighted combination fusion feature space is constructed, and learnable fusion weights satisfying non-negative normalization constraints are introduced. This allows for adaptive fusion of the feature representations of multiple stage adapters, with the goal of maximizing the pairwise distance between category prototypes within the fusion space. Through these techniques, the collaborative reconstruction of cross-stage related feature representations is achieved without compromising the stability of the original stage knowledge. This results in higher similarity categories obtaining greater separation distances in the unified fusion space, thereby improving the separability between confused categories and enhancing the fine-grained distinguishability during the inference stage.

[0020] (3) To address the lack of a dedicated reasoning mechanism for potentially confusing categories in existing technologies, which typically involve direct classification decisions within a unified feature space and difficulty in further distinguishing highly similar categories, this invention, after completing the modeling of confusion relationships and the construction of the fusion space, performs reasoning based on category prototypes. For categories belonging to confused connected branches, refined prototype distance calculation and discrimination are prioritized within their corresponding fusion feature space. For isolated categories not involved in cross-stage confusion relationships, their original stage adaptation modules are retained for feature extraction and classification decisions. Through this branch differentiation strategy, the model concentrates computational resources on easily confused category regions while maintaining controllable overall computational overhead, achieving targeted enhanced discrimination and improving overall classification accuracy and reasoning stability.

[0021] (4) To address the challenge of balancing discriminative ability with training cost control in incremental learning, this invention constructs a complete processing flow at the overall framework level, including efficient training of stage adapter parameters, cross-stage confusion graph modeling, connected component-guided fusion enhancement, and prototype-based hierarchical reasoning. This framework requires only training a small number of newly added adapter module parameters at each stage, and performs relationship analysis through category prototypes during the confusion modeling stage, eliminating the need to backtrack to large-scale historical data. This allows for refined modeling and differentiation of cross-stage confusion categories while maintaining efficient parameter fine-tuning and anti-forgetting capabilities, thereby improving the system's stability and practicality in scenarios with continuously expanding long-sequence categories. Attached Figure Description

[0022] Figure 1 This is a framework diagram of the model training stage in an embodiment of the present invention; Figure 2 This is a framework diagram of the model inference stage in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] Example 1: This embodiment provides a category incremental classification method based on disjoint-set graph modeling and fusion space. The overall technical framework is as follows: Figure 1 and Figure 2 As shown, it includes the following steps: Step S1: Construct an incremental learning task environment, load the pre-trained visual backbone network and freeze its parameters, and configure an independent adaptation module for each incremental stage in the visual backbone network. In one specific embodiment, a class incremental learning task environment is first constructed, and the class incremental learning process is set to include a total number of incremental stages. The datasets corresponding to each incremental stage are as follows: The category sets at different incremental stages do not overlap, satisfying... ,in This ensures that only new categories are introduced at each stage, thus conforming to the task setting of incremental class learning.

[0025] Dataset for each incremental stage This includes the set of training samples for the newly added categories in this incremental phase and their corresponding set of category labels, specifically represented as follows: in, Indicates the first The first incremental stage There are one input sample, and the input sample is image data; This indicates the category label corresponding to the input sample; Indicates the first The total number of samples in each incremental phase; The dataset for the incremental stage is used to supervise the training of the adaptation module for the corresponding incremental stage. This allows the adaptation module to learn the feature representation of the newly added categories in this stage while the parameters of the visual backbone network are frozen. The dataset is then used to calculate the category prototypes for each category in this stage after training is completed.

[0026] Load the pre-trained Vision Transformer visual backbone network model. This model includes... L Each Transformer encoder layer consists of a multi-head self-attention module and a feedforward multilayer perceptron module. Pre-trained weights are obtained from a large-scale image dataset (e.g., ImageNet-1K). After loading, all parameters of the visual backbone network are frozen, and no gradient updates are performed during subsequent incremental training stages. By freezing the backbone network parameters, the general visual representation capability of the pre-trained model can be fully preserved, avoiding interference with learned knowledge during new category learning, thus reducing the risk of catastrophic forgetting, and significantly reducing the size of training parameters. An adaptation module is inserted into each feature transformation layer of the visual backbone network. , Each incremental stage corresponds to an independent adaptation module, which adopts a bottleneck structure, including a dimensionality reduction mapping matrix. Nonlinear activation functions and up-dimensional mapping matrices The adapter module is inserted into the feedforward bypass of each layer of the visual backbone network in the form of residual connections. The output of the feedforward network is corrected as follows: in, This represents the input feature representation of the corresponding layer in the visual backbone network. The output of the feedforward network after the integration and adaptation module; This represents the feedforward neural network module of this layer in the visual backbone network; Represents the linear rectification activation function; Initialize the corresponding adaptation module parameters for each incremental stage. The parameters of the adaptation modules in each incremental stage are independent and not shared, and are used to learn the feature representation of the corresponding stage category at different stages.

[0027] Step S2: According to the preset incremental stage order, the corresponding adaptation module of the incremental stage is trained sequentially using the newly added category data of the corresponding incremental stage. After the training is completed, its parameters are fixed, and the feature representation of each category of the incremental stage is extracted based on the trained visual backbone network and the corresponding adaptation module to construct the category prototype set of the incremental stage. In one specific embodiment, for the first In the first incremental stage, the parameters of the visual backbone network are fixed, and only the first... The corresponding adaptation module for each incremental stage parameters Set it to a trainable state; by updating only the parameters of the current stage's adaptation module, the model can gain the ability to express new categories while maintaining the stability of the existing stage's feature representation, thus avoiding mutual interference between parameters of different stages. Use the dataset of this incremental stage. The input consists of a visual backbone network and an adaptation module. The feature extraction model is composed of a feature representation to obtain the sample feature representation; a classification prediction function is constructed based on the feature representation, and the prediction output is obtained through a linear classifier; the linear classifier is composed of a weight matrix and a bias vector, and its output is the result of a linear transformation of the feature vector, which is used to calculate the predicted distribution of the newly added categories in the current stage.

[0028] A supervised loss function is constructed based on the true class labels and predicted outputs in the dataset. The loss function adopts cross-entropy loss. The gradient descent optimization method is used to minimize the loss function, and only the parameters of the adaptation module and the classifier parameters of the corresponding stage are updated until the preset convergence condition is met or the maximum number of training rounds is reached.

[0029] In the After each incremental training phase is completed, the adaptation module is fixed. After setting the parameters, remove or freeze the classifier for the corresponding stage, retaining only the frozen visual backbone network and the first stage. Each incremental stage adaptation module serves as a feature extraction model; The dataset corresponding to the incremental stage The input samples are re-inputted into the feature extraction model to obtain the feature representation of each sample: in, Indicates the first In the first incremental stage Feature representation of each sample; This indicates that the input sample is frozen and connected to the visual backbone network. The feature mapping function after processing by the incremental stage adaptation module; Indicates the first The first incremental stage One input sample; Indicates the first The corresponding adaptation modules for each incremental stage; For the Any category in the incremental phase Select the category labels that meet the criteria. The feature representations of all samples are then aggregated by mean to obtain the category prototype for that category: in, Indicates the first Categories in each incremental stage The category prototype; Indicates the first Categories in each incremental stage The number of samples; Indicates the first The total number of samples in each incremental phase; This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise; Indicates the first In the first incremental stage The category labels of each sample; Indicates the first The category set in each incremental stage; using the category feature mean as the category prototype can reduce the impact of individual sample noise on the category representation to a certain extent, making the category center more stable, thus providing a more robust basic representation for subsequent category similarity calculation and graph structure modeling.

[0030] The first The category prototypes corresponding to all categories in each incremental stage constitute the category prototype set for that incremental stage: in, Indicates the first A set of category prototypes for each incremental stage.

[0031] Step S3: After completing all incremental training phases, summarize the class prototype sets of all incremental phases to form a global class prototype set; Step S4: Based on the global category prototype set, calculate the similarity relationship between category prototypes from different incremental stages. When the preset confusion judgment condition is met, establish the confusion association between categories and construct a cross-stage category confusion relationship set. In one specific embodiment, after completing training for all incremental stages and obtaining a global set of category prototypes, pairwise similarity calculations are performed on the category prototypes between different incremental stages to uncover potential highly similar category pairs across stages, thereby constructing a set of cross-stage category confusion relationships. This step unifies the category representations, originally scattered across different stages and trained independently, into a single analytical framework, providing foundational data support for subsequent graph structure modeling and connected component identification.

[0032] Suppose that the global set of category prototypes contains all category prototypes from each incremental stage. For any two category prototypes... and When its incremental stage satisfies At that time, prototype similarity is calculated. The similarity between category prototypes is measured using cosine similarity, and the calculation formula is: in, Indicates category With category Prototype similarity; Indicates the first Categories in each incremental stage The category prototype; Indicates the first Categories in each incremental stage The category prototype; The L2 norm of a vector; Set a preset confusion threshold When satisfied When determining the category With category Confusional relationships exist in the feature space; confusional edges are established between categories, denoted as... Otherwise, record as ;in, Indicates category With category The confusion between the indicator variables, when A '-' indicates that a confusing association is established; otherwise, a confusing association is not established. All satisfied and The category pairs constitute a set of cross-stage category confusion relationships: in, This represents a set of cross-stage category confusion relationships; This indicates a confusion association pair that is determined to have a confusion relationship between different incremental stages.

[0033] A threshold control mechanism can be used to filter out class pairs with significant similarity in the feature space, thereby avoiding redundant connections introduced by low-similarity classes and ensuring the sparsity and effectiveness of the subsequent graph structure. The threshold value can be adjusted based on validation set performance or empirical settings to balance the sensitivity and stability of confusion detection.

[0034] Step S5: Based on the set of category confusion relationships, construct a cross-stage category confusion graph with each category as a node and confusion association as an edge, and use a disjoint-set data structure to perform connectivity merging processing on the confusion graph to obtain several category confusion connected branches; In one specific embodiment, after obtaining the cross-stage category confusion relationship set, a cross-stage category confusion graph is constructed using each category as a graph node and the confusion associations between categories as edges. A disjoint-set data structure is then used to perform connectivity merging on the graph structure, thereby identifying several category confusion connectivity branches. Through this process, pairs of categories with confusion relationships can be further expanded into structured category clusters, avoiding a focus solely on local pairwise relationships and achieving a unified characterization of the global confusion structure. The categories in the global category prototype set are used as nodes in the graph, and the category confusion relationship set... Each pair of confusing associations As edges of the graph; obtain a cross-stage category confusion graph. , is represented as: in, The set of nodes in the graph, including category nodes for all incremental stages. ; The set of edges of the graph, including the set of category confusion relations. E All confused association pairs; initialize the disjoint-set data structure, setting each category as an independent set, with the initial state of the disjoint-set being that each category node is a separate connected component; for category nodes The corresponding initial set is: in, Represents category nodes The parent node, Represents category nodes The depth of the tree; Traversing the confusion graph Each edge in For each pair of obfuscated associations, perform the following operations: For each pair of category nodes and Use the Find operation of the disjoint-set data structure to search for each element. and The root node, if If so, a Union operation is performed to merge the sets containing the two types of nodes, and the corresponding sets are updated according to the rank information. and Parameters. If the ranks of the two trees are different, the tree with the smaller rank is connected to the tree with the larger rank; if the ranks are the same, either tree is chosen as the new root node, and its rank is incremented by one. After processing all edges, a disjoint-set data structure containing several connected components is obtained, where each connected component represents a set of categories connected in the graph through confusion relations. By sequentially performing the above merging operation on all confused edges, category nodes with direct or indirect confusion relations can be gradually merged into the same set. This process not only considers directly similar category pairs but can also identify indirectly connected category sets through transitivity relations. For example, when categories c1 and c2 have a confusion relationship, and categories c2 and c3 have a confusion relationship, even if c1 and c3 do not directly meet the similarity threshold condition, they can still be classified into the same confused connected component through connectivity analysis, thereby achieving the overall identification of complex confusion structures. Through each independent set in the disjoint-set data structure, several confused connected components are divided, each composed of multiple category nodes connected to each other through confusion relations; the confused connected components are filtered to select those containing at least two category nodes, which are the final category-confused connected components.

[0035] Step S6: For the confused connected branches, select the corresponding adaptation module to construct the fusion feature space according to the incremental stage to which each category belongs in the branch, and enhance the distinguishability between categories in the confused connected branches by optimizing the fusion weights, and obtain the fusion feature space model corresponding to each confused connected branch. In one specific embodiment, for each filtered confused connected component ,in, The set of nodes in the graph; based on the nodes of each category The incremental stage to which it belongs Select the corresponding pre-trained and fixed-parameter adaptation module. As a module for feature extraction in this category; Based on the confusion of all categories of nodes within the connected component Corresponding to the incremental phase, an adaptation module is issued. Define the fusion feature space of confused connected components. A weighted combination of features for the selected adaptation modules: in, Indicates confusing connected components The fusion feature space; Indicates confusing connected components The number of nodes in each category; Indicates the corresponding adaptation module The learnable fusion weight coefficients are used to adjust the contribution of features in each incremental stage to the fusion space; Indicates the first Adaptation modules for each incremental stage; Initialize fusion weight coefficients And satisfy the constraints: This will obfuscate the category prototypes of various categories within the connected component. Mapping to the fused feature space The fused feature prototype is obtained. ; In the fusion feature space In the middle, calculate the feature prototypes of all categories within the confusing connected component. Pairwise Euclidean distance between them: in, and , Indicates Euclidean distance; Based on the feature prototype The pairwise Euclidean distance between them defines the optimization objective of confusing connected components. : To maximize the optimization objective To achieve the objective, the gradient descent method is used to optimize the fusion weight coefficients. Iterate and update until convergence or the preset maximum number of iterations is reached, and save the optimized fusion weight coefficients and the corresponding fusion feature space description information to form confused connected components. Fusion feature space model .

[0036] Through the above-mentioned process of constructing the fusion space and optimizing the weights, while maintaining the stability of knowledge at the original stage, the gaps between highly similar categories across stages can be specifically widened, improving the separability between confused categories and providing more discriminative feature representations for subsequent refined prototype reasoning.

[0037] For connected components that contain only a single category node (i.e., isolated categories without cross-stage confusion relationships), there is no need to construct a fused feature space; the adapter of its original stage can be directly retained as the feature extractor.

[0038] Step S7: In the inference phase, for the sample to be classified, the sample features are extracted through the adaptation modules of each incremental phase and matched with the global category prototype set to obtain the preliminary classification result. In one specific embodiment, the sample to be classified x The frozen visual backbone network is input and forward-propagated through the adaptation modules of each incremental stage to obtain the sample features of each stage: in, Indicates the sample to be classified After freezing the visual backbone network and the first Incremental Stage Adaptation Module Features of the processed samples; Indicates the total number of incremental stages; For each incremental stage Calculate sample features Cosine similarity with each category prototype in the incremental stage category prototype set The similarity calculation results from all stages are summarized to generate a global similarity ranking list, and the top similarity scores are selected. Each category is used as a candidate category set. : Each candidate category contains a category identifier. Belonging to the incremental stage and corresponding similarity scores ; Indicates sample With incremental stage category Cosine similarity; The candidate category set This serves as a preliminary classification result.

[0039] Step S8: When there are candidate categories belonging to the same confused connected component and from different incremental stages in the preliminary classification results, call the corresponding fusion feature space model to further distinguish the candidate categories and output the final classification result; if there are no such categories, directly output the preliminary classification result as the final classification result. In one specific embodiment, the preliminary classification results are refined to obtain the final classification result. The affiliation relationship between the top-1 candidate category and other candidate categories in the preliminary classification results is examined; if at least one candidate category exists... The second level of fine-tuning is triggered when the following two conditions are met: (a) It belongs to the same confusing connected component as the top-1 candidate category; (b) The top-1 candidate category comes from a different incremental phase; When the second level of fine-tuning is triggered, the obfuscated connectivity branch is invoked. Fusion feature space model The samples to be classified The features are mapped to the fused feature space to obtain the samples to be classified. Enhanced feature representation in the fusion feature space; this enhanced feature representation effectively enhances the discriminativeness of cross-stage similar categories by fusing the features of all relevant incremental stage adaptation modules in the confusion branch, and alleviates the problem of misjudgment of easily confused categories in the initial classification.

[0040] In the fusion feature space In the middle, calculate the samples to be classified Enhanced feature representation and obfuscated connected components Feature prototypes of all categories Cosine similarity; The category with the highest cosine similarity is used as the sample to be classified. The final predicted category; If the preliminary classification results do not trigger the second-level refinement condition, the top-1 candidate category is directly output as the final predicted category without additional calculation. This mechanism ensures high efficiency in inference for most samples, while significantly improving classification accuracy and model robustness by fusing feature space refinement in the presence of potential cross-stage confusion.

[0041] Example 2: This embodiment provides a category incremental classification system based on disjoint-set graph modeling and fusion space, including: The data management module is used to build an environment for incremental learning tasks, manage datasets for each incremental stage, and provide training samples and corresponding category labels. The feature extraction module includes a pre-trained visual backbone network and an independent adaptation module for each incremental stage, used to extract feature representations of each stage category; The category prototype construction module generates a set of category prototypes for each incremental stage by aggregating the mean of features of samples of the same category. The confusion modeling module is used to calculate cross-stage category prototype similarity relationships, identify potential confusion categories, and construct a set of category confusion relationships. The connected component analysis module constructs a confusion graph from categories and confusion relationships, and uses a disjoint-set data structure to identify confused connected components; The fusion feature space module performs weighted fusion of features of categories within confusing connected components, enhances the ability to distinguish similar categories across stages, and generates a fusion feature space model. The inference module extracts features from the samples to be classified through the adaptation modules of each incremental stage, and matches them with the global category prototype to generate preliminary classification results. The precision judgment module calls the fusion feature space model to perform secondary discrimination on samples with cross-stage confusion and outputs the final classification result.

[0042] This system achieves accurate differentiation of similar categories across stages through structured confusion modeling and feature fusion, while maintaining low training costs and resistance to forgetting.

[0043] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0044] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for incremental category classification based on disjoint-set graph modeling and fusion space, characterized in that, Includes the following steps: Construct an incremental learning task environment, load a pre-trained visual backbone network and freeze its parameters, and configure an independent adaptation module for each incremental stage in the visual backbone network. According to the preset incremental stage order, the corresponding adaptation module of the incremental stage is trained sequentially using the newly added category data of the corresponding incremental stage. After the training is completed, its parameters are fixed, and the feature representation of each category of the incremental stage is extracted based on the trained visual backbone network and the corresponding adaptation module to construct the category prototype set of the incremental stage. After completing all incremental training phases, the class prototype sets of all incremental phases are aggregated to form a global class prototype set. Based on the global category prototype set, the similarity relationship between category prototypes from different incremental stages is calculated. When the preset confusion judgment condition is met, the confusion association between categories is established, and a cross-stage category confusion relationship set is constructed. Based on the set of category confusion relationships, a cross-stage category confusion graph is constructed with each category as a node and the confusion association as an edge. The confusion graph is then processed by a disjoint-set data structure to obtain several category confusion connected branches. For the confused connected branches, based on the incremental stage to which each category belongs within the branch, the corresponding adaptation module is selected to construct the fusion feature space, and the ability to distinguish between categories within the confused connected branches is enhanced by optimizing the fusion weights, thereby obtaining the fusion feature space model corresponding to each confused connected branch. During the inference phase, for the sample to be classified, the sample features are extracted through the adaptation modules of each incremental phase and matched with the global category prototype set to obtain preliminary classification results. When the preliminary classification results contain candidate categories belonging to the same confused connected component but from different incremental stages, the corresponding fusion feature space model is invoked to further discriminate the candidate categories and output the final classification result. If it does not exist, the preliminary classification result will be output directly as the final classification result.

2. The incremental classification method for categories based on disjoint-set graph modeling and fusion space according to claim 1, characterized in that, The constructed incremental learning task environment specifically includes: The incremental learning task is defined as having a total number of incremental stages. The datasets corresponding to each incremental stage are as follows: Where the datasets in any two different incremental stages do not overlap in categories, satisfying ,in ; Dataset for each incremental stage This includes the set of training samples for the newly added categories in this incremental phase and their corresponding set of category labels, specifically represented as follows: in, Indicates the first The first incremental stage One input sample, wherein the input sample is image data; This represents the category label corresponding to the input sample; Indicates the first The total number of samples in each incremental phase; The dataset for the incremental stage is used to supervise the training of the adaptation module for the corresponding incremental stage, so that the adaptation module learns the feature representation of the newly added category in the stage under the condition of freezing the visual backbone network parameters, and is used to calculate the category prototype of each category in the stage after training is completed.

3. The incremental classification method for categories based on disjoint-set graph modeling and fusion space according to claim 1, characterized in that, The process of loading the pre-trained visual backbone network and freezing its parameters, and configuring an independent adaptation module for each incremental stage in the visual backbone network, specifically includes: Load the pre-trained visual backbone network and freeze all its network parameters so that the weights are not updated during each incremental training phase. An adapter module is inserted into each feature transformation layer of the visual backbone network. , Each incremental stage corresponds to an independent adaptation module, which adopts a bottleneck structure and includes a dimensionality reduction mapping matrix. Nonlinear activation functions and up-dimensional mapping matrices ; The adaptation module is inserted into the feedforward network bypass of each layer of the visual backbone network in the form of residual connections, and the output of the feedforward network is corrected as follows: in, This represents the input feature representation of the corresponding layer in the visual backbone network. The output of the feedforward network after the integration and adaptation module; This represents the feedforward neural network module of this layer in the visual backbone network; Represents the linear rectification activation function; Initialize the corresponding adaptation module parameters for each incremental stage. The parameters of the adaptation modules in each incremental stage are independent and not shared, and are used to learn the feature representation of the corresponding stage category at different stages.

4. The incremental classification method for categories based on disjoint-set graph modeling and fusion space according to claim 1, characterized in that, The step of training the adaptation module corresponding to the incremental stage according to the preset incremental stage sequence, using the newly added category data of the corresponding incremental stage, specifically includes: For the In the first incremental stage, the parameters of the visual backbone network are fixed, and only the first... The corresponding adaptation module for each incremental stage parameters Set to trainable state; The dataset of this incremental phase The input consists of a visual backbone network and an adaptation module. The feature extraction model is composed of these components to obtain the feature representation of the samples. A classification prediction function is constructed based on the feature representation, and the prediction output is obtained through a linear classifier. A supervised loss function is constructed based on the true class labels and predicted outputs in the dataset, and the loss function adopts cross-entropy loss; The loss function is minimized using the gradient descent optimization method, updating only the parameters of the adaptation module and the classifier parameters of the corresponding stage, until the preset convergence condition is met or the maximum number of training rounds is reached.

5. The incremental classification method for categories based on disjoint-set graph modeling and fusion space according to claim 1, characterized in that, The step of extracting feature representations of each category in the incremental stage based on the trained visual backbone network and corresponding adaptation modules, and constructing a category prototype set for the incremental stage, specifically includes: In the After each incremental training phase is completed, the adaptation module is fixed. After setting the parameters, remove or freeze the classifier for the corresponding stage, retaining only the frozen visual backbone network and the first stage. Each incremental stage adaptation module serves as a feature extraction model; The dataset corresponding to the incremental stage The input samples are re-inputted into the feature extraction model to obtain the feature representation of each sample: in, Indicates the first In the first incremental stage Feature representation of each sample; This indicates that the input sample is frozen and connected to the visual backbone network. The feature mapping function after processing by the incremental stage adaptation module; Indicates the first The first incremental stage One input sample; Indicates the first The corresponding adaptation modules for each incremental stage; For the Any category in the incremental phase Select the category labels that meet the criteria. The feature representations of all samples are then aggregated by mean to obtain the category prototype for that category: in, Indicates the first Categories in each incremental stage The category prototype; Indicates the first Categories in each incremental stage The number of samples; Indicates the first The total number of samples in each incremental phase; This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise; Indicates the first In the first incremental stage The category labels of each sample; Indicates the first A set of categories for each incremental stage; The first The category prototypes corresponding to all categories in each incremental stage constitute the category prototype set for that incremental stage: in, Indicates the first A set of category prototypes for each incremental stage.

6. The incremental classification method for categories based on disjoint-set graph modeling and fusion space according to claim 1, characterized in that, Based on the global category prototype set, the similarity relationship between category prototypes from different incremental stages is calculated. When a preset confusion judgment condition is met, confusion associations between categories are established, and a cross-stage category confusion relationship set is constructed, specifically including: For any two categories in the global category prototype set, arbitrarily select two category prototypes from different incremental phases from the global category prototype set. and When satisfied Then, calculate the prototype similarity between the two: in, Indicates category With category Prototype similarity; Indicates the first Categories in each incremental stage The category prototype; Indicates the first Categories in each incremental stage The category prototype; The L2 norm of a vector; Set a preset confusion threshold When satisfied When determining the category With category Confusional relationships exist in the feature space; confusional edges are established between categories, denoted as... Otherwise, record as ;in, Indicates category With category The confusion between the indicator variables, when A '-' indicates that a confusing association is established; otherwise, a confusing association is not established. All satisfied and The category pairs constitute a set of cross-stage category confusion relationships: in, This represents a set of cross-stage category confusion relationships; This indicates a confusion association pair that is determined to have a confusion relationship between different incremental stages.

7. The incremental classification method for categories based on disjoint-set graph modeling and fusion space according to claim 1, characterized in that, Based on the set of category confusion relationships, a cross-stage category confusion graph is constructed with each category as a node and the confusion association as an edge. A disjoint-set data structure is then used to perform connectivity merging on the confusion graph, resulting in several category confusion connectivity branches, specifically including: Using categories from the global category prototype set as nodes in the graph, a set of category confusion relationships is created. Each pair of confusing associations As edges of the graph; obtain a cross-stage category confusion graph. , is represented as: in, The set of nodes in the graph, including category nodes for all incremental stages. ; The set of edges of the graph, including the set of category confusion relations. E All obfuscated association pairs; Initialize the disjoint-set data structure, setting each category as an independent set. The initial state of the disjoint-set is that each category node is a separate connected component. For category nodes The corresponding initial set is: in, Represents category nodes The parent node, Represents category nodes The depth of the tree; Traversing the confusion graph Each edge in For each pair of obfuscated associations, perform the following operations: For each pair of category nodes and Use the Find operation of the disjoint-set data structure to search for each element. and The root node, if Then, a Union operation is performed to merge the sets containing nodes of the two categories, and the corresponding updates are made. and information: After processing all edges, a disjoint set containing several connected components is obtained, where each connected component represents a set of categories connected in the graph through confusion relations; Each independent set in the disjoint-set data structure is divided into several obfuscated connected branches. Each obfuscated connected branch consists of multiple category nodes, which are interconnected through obfuscation relationships. The confused connected branches are filtered to select those containing at least two categories of nodes, which are then used as the final category-confused connected branches.

8. The incremental classification method for categories based on disjoint-set graph modeling and fusion space according to claim 1, characterized in that, For the confused connected component, based on the incremental stage to which each category belongs within the component, a corresponding adaptation module is selected to construct a fusion feature space. Furthermore, the ability to distinguish between categories within the confused connected component is enhanced by optimizing the fusion weights, thereby obtaining a fusion feature space model corresponding to each confused connected component. Specifically, this includes: For each filtered confusing connected component ,in, Let be the set of nodes in the graph; Based on each category node The incremental stage to which it belongs Select the corresponding pre-trained and fixed-parameter adaptation module. As a module for feature extraction in this category; Based on the confusion of all categories of nodes within the connected component Corresponding to the incremental phase, an adaptation module is issued. Define the fusion feature space of confused connected components. A weighted combination of features for the selected adaptation modules: in, Indicates confusing connected components The fusion feature space; Indicates confusing connected components The number of nodes in each category; Indicates the corresponding adaptation module The learnable fusion weight coefficients are used to adjust the contribution of features in each incremental stage to the fusion space; Indicates the first Adaptation modules for each incremental stage; Initialize fusion weight coefficients And satisfy the constraints: This will obfuscate the category prototypes of various categories within the connected component. Mapping to the fused feature space The fused feature prototype is obtained. ; In the fusion feature space In the middle, calculate the feature prototypes of all categories within the confusing connected component. Pairwise Euclidean distance between them: in, and , Indicates Euclidean distance; Based on the feature prototype The pairwise Euclidean distance between them defines the optimization objective of confusing connected components. : To maximize the optimization objective To achieve the objective, the gradient descent method is used to optimize the fusion weight coefficients. Iterate and update until convergence or the preset maximum number of iterations is reached, and save the optimized fusion weight coefficients and the corresponding fusion feature space description information to form confused connected components. Fusion feature space model .

9. The incremental classification method for categories based on disjoint-set graph modeling and fusion space according to claim 1, characterized in that, In the inference phase, for the sample to be classified, sample features are extracted through the adaptation modules of each incremental phase and matched with the global category prototype set to obtain preliminary classification results, specifically including: Samples to be classified x The frozen visual backbone network is input and forward-propagated through the adaptation modules of each incremental stage to obtain the sample features of each stage: in, Indicates the sample to be classified After freezing the visual backbone network and the first Incremental Stage Adaptation Module Features of the processed samples; Indicates the total number of incremental stages; For each incremental stage Calculate sample features Cosine similarity with each category prototype in the incremental stage category prototype set ; The similarity calculation results from all stages are aggregated to generate a global similarity ranking list, and the top similarity scores are selected. Each category is used as a candidate category set. : Each candidate category contains a category identifier. Belonging to the incremental stage and corresponding similarity scores ; Indicates sample With incremental stage category Cosine similarity; Candidate category set This serves as a preliminary classification result.

10. The incremental classification method for categories based on disjoint-set graph modeling and fusion space according to claim 1, characterized in that, When there are candidate categories in the preliminary classification results that belong to the same confused connected component but come from different incremental stages, the corresponding fusion feature space model is called to re-discriminate the candidate categories and output the final classification result. If it does not exist, the preliminary classification result will be directly output as the final classification result, specifically including: Examine the attribution relationship between the top-1 candidate category and other candidate categories in the preliminary classification results; if there is at least one candidate category... The second level of fine-tuning is triggered when the following two conditions are met: (a) It belongs to the same confusing connected component as the top-1 candidate category; (b) The top-1 candidate category comes from a different incremental phase; When the second level of fine-tuning is triggered, the obfuscated connectivity branch is invoked. Fusion feature space model The samples to be classified The features are mapped to the fused feature space to obtain the samples to be classified. Enhanced feature representation in the fused feature space; In the fusion feature space In the middle, calculate the samples to be classified Enhanced feature representation and obfuscated connected components Feature prototypes of all categories Cosine similarity; The category with the highest cosine similarity is used as the sample to be classified. The final predicted category; If the second-level fine-tuning condition is not triggered, the top-1 candidate category of the preliminary classification result will be directly output as the final predicted category.

Images