Method for relieving catastrophic forgetting of incremental target detection based on incremental relationship consistency
By co-designing the twin contrast learning module and the boundary-aware structure preservation module, the catastrophic forgetting problem in incremental object detection models is solved, and the detection performance of old categories is stably retained when learning new categories, thereby improving the model's cross-task adaptability and generalization ability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF TECH
- Filing Date
- 2025-12-12
- Publication Date
- 2026-04-21
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing incremental object detection models suffer from significant forgetting of previously learned knowledge of old categories when learning new categories, marked performance degradation across tasks, serious knowledge shift issues, unstable feature structures, insufficient utilization of cross-task relationships, and severe performance degradation in multi-step incremental steps, failing to meet the needs of dynamic applications.
The Siamese Contrast Learning Module (SCRL) and the Boundary Aware Structure Preservation Module (BASP) work together to learn general cross-task relations through the Siamese Contrast Learning Module, while the Boundary Aware Structure Preservation Module maintains the class center and decision boundary. Combined with the total loss optimization objective, the model can stably retain the detection performance of the old class when learning new classes.
It mitigates catastrophic forgetting, stabilizes feature structure, improves cross-task adaptability, enhances the model's generalization ability in single-step and multi-step incremental scenarios, simplifies the training process, and promotes the practical application of incremental detection technology.
Smart Images
Figure CN121904545A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of catastrophic forgetting technology, specifically a method for mitigating catastrophic forgetting in incremental target detection based on incremental relationship consistency. Background Technology
[0002] Catastrophic forgetting mitigation is a core challenge in incremental learning, and existing research can be mainly categorized into three types: regularization methods (such as LwM, EWC, LwF), distillation methods (such as PODNet, Dytox, DER), and replay methods (such as EEiL, iCarL, LUCIR, SDDGR).
[0003] Regularization methods impose constraints on model parameters, penalizing the update magnitude of old parameters, thus enabling the model to retain existing knowledge when fine-tuning with new data.
[0004] Its limitations lie in the insufficient constraint strength when facing multi-task or significantly different task scenarios, which can still lead to forgetting, and the potential bias in parameter importance estimation. The distillation-based method generates soft targets from the old model, and the new model mimics the old model's output by minimizing the cross-entropy loss between the predicted probability and the soft target, thus achieving the transfer of old knowledge to new tasks while avoiding forgetting.
[0005] However, simple output or feature distillation can easily lead to knowledge shift issues. Replay-based methods store partial samples from previous tasks, and then combine new data with the stored samples during subsequent training to address the problem of missing old data. These methods require storing a small amount of old data or additional training to generate the model, which may bring privacy and storage pressures, and their performance is highly sensitive to memory bank management strategies.
[0006] Meanwhile, existing methods do not fully utilize intra- and inter-category relational knowledge, relying only on task-specific semantic features, resulting in poor model adaptability and more significant performance degradation in multi-step incremental learning scenarios.
[0007] Specifically as follows:
[0008] 1. Severe catastrophic amnesia:
[0009] Existing incremental object detection models show significant forgetting of previously learned knowledge of old categories when learning new categories, resulting in a marked decline in cross-task performance.
[0010] 2. Knowledge bias problem:
[0011] Direct feature alignment distillation cannot adapt to semantic heterogeneity between tasks, resulting in distorted feature representations for old tasks and insufficient model robustness.
[0012] 3. Unstable feature structure:
[0013] Without effective constraints on feature distribution, new category features can easily intrude into the feature space of old categories, disrupting category boundaries.
[0014] 4. Insufficient utilization of cross-task relationships:
[0015] Existing models rely on task-specific semantic features and have not fully learned cross-task general intra-class and inter-class relationships, resulting in poor adaptability.
[0016] 5. Multi-step incremental performance degradation:
[0017] In multi-step incremental learning scenarios, the performance of existing methods suffers from severe cumulative degradation, failing to meet the needs of dynamic applications.
[0018] Therefore, a new solution is needed to address the above problems. Summary of the Invention
[0019] The purpose of this invention is to provide a method for mitigating catastrophic forgetting in incremental target detection based on incremental relationship consistency, so as to solve the technical problems mentioned in the background art.
[0020] To achieve the above objectives, the present invention provides the following technical solution: a method for mitigating catastrophic forgetting in incremental target detection based on incremental relationship consistency, comprising at least the following steps:
[0021] S1: Perform data processing and initialization;
[0022] S2: Construct a Siamese contrastive relation learning module, namely the SCRL module. The query image and support images are respectively input into a weight-shared ResNet-50 backbone network, and the output is the basic features of the query. and supporting basic features Where c is the number of categories and d is the feature dimension; for A multi-head self-attention mechanism is applied to enhance the pixel correlation within the feature map and suppress background noise; the supporting features are downsampled to a fixed size of 7×7 through pooling operations to unify the dimensions and reduce the computational complexity of subsequent operations.
[0023] S3: Construct a boundary-aware structure preservation module, namely the BASP module, which maintains the class center and decision boundary through a constraint loss function;
[0024] S4: Use total loss to optimize the overall objective;
[0025] S5: Conduct model evaluation and testing.
[0026] Furthermore, S1 includes at least the following steps:
[0027] Before each incremental task is started, for the t-th incremental task, the ABR (AugmentedBoxReplay) method is used to select representative target regions from the new task data to crop images and form a candidate set of supporting images. Then, it is merged with the historical task cache to obtain a complete support set B covering all old and new categories.
[0028] Next, the model parameters are initialized:
[0029] The first task initializes the ResNet-50 backbone network parameters θ and the Faster R-CNN classifier / regressor head parameters φ.
[0030] When it is a subsequent incremental task, it inherits the parameters θt-1 and φt-1 after training the previous task, and initializes the parameters required by the BASP module, such as class center and class boundary radius r.
[0031] Finally, the input data is divided, and the complete image to be detected in the current task is used as the query image (for final detection), while the target cropped image randomly selected from the memory cache B is used as the support image (to provide cross-task feature reference).
[0032] Furthermore, the application of the SCRL module includes at least the following steps:
[0033] A randomly initialized task code T is introduced, and the features of the query set interact with the new task code to learn new category specificities while preserving general relationships;
[0034] Cross-attention calculation is represented as:
[0035]
[0036] Where W1, W2, and W3 are learnable attention parameters, and A T Support features after interaction (only general relational semantics are retained). For query matrix; The key matrix;
[0037] Output features are processed using a single-head attention mechanism. The similarity between the query features and supporting features is calculated, and features irrelevant to the supported categories are filtered out. The matching coefficient is calculated as follows:
[0038]
[0039] Where W is The linear projection shared with Q, This represents the activation function. Represents element-level multiplication. Represents the query matrix;
[0040] The attention module is used to accurately identify the semantic relevance between query features and supporting features in object detection.
[0041] The fused features are then used for detection, following an inclusive distillation loss, and distillation is performed on the output to enhance the category knowledge of the old task:
[0042]
[0043] in, It is a normalization factor;
[0044] For the old task category The distillation loss at the current time step represents the current time step. The target category and related categories (from) -1 to The relationship between the probabilities of )
[0045] For all categories The loss calculation involves comparing the class information from the previous time step with the class probability of the current time step.
[0046] Furthermore, the application of the BASP module includes at least the following steps:
[0047] The first task training loss function is proposed in the classification head of the detector, denoted as:
[0048]
[0049] in, An index representing all categories; Representative category The sample set; Indicates the size of the set; Indicates sample eigenvectors; It is a category The central eigenvector; Represents a constant; This represents the temperature parameter, used to control the smoothness of the loss function.
[0050] After completing the training for the first task, estimate the boundary of the feature distribution for each category to determine the reserved boundary for subsequent tasks;
[0051] To simplify the analysis, the feature distribution of each category is modeled as follows: , The mean, For variance;
[0052] The boundary should be defined to include 90% of the distribution within the boundary; the corresponding hyperparameter is... (Right now ),and The time-based model has the best detection performance;
[0053] The boundary can be calculated as follows:
[0054]
[0055]
[0056] in It's a hyperparameter. The inverse cumulative distribution function of the standard normal distribution;
[0057] Used to calculate the upper and lower boundaries of the category feature distribution;
[0058] Through linear transformation, it is possible to determine Distribution median probability The corresponding left and right coordinate values;
[0059] Calculate the center threshold of the confidence interval based on the margin:
[0060]
[0061] The center threshold defines the decision boundary of the old task in the feature space, preserves the core region of the old task's feature space, and creates a buffer for subsequent tasks.
[0062] Calculate all boundary thresholds Then, the maximum historical boundary is defined as Used to dynamically adjust the marginal parameters in the loss function:
[0063]
[0064] Constraints Ensure there is a boundary between the center of the new task category and the maximum boundary of the old task. The minimum safe distance, this constraint ensures that there is a minimum safe distance between the center of the new task category and the maximum boundary of the old task. The minimum safe distance is used to prevent new category features from interfering with the distribution of old task features.
[0065] Furthermore, the total loss includes regression loss, classification loss from BASP, and inclusive distillation loss. .
[0066] Furthermore, S5 includes at least the following steps:
[0067] The test set images of the Pascal VOC and MS COCO benchmark datasets are input into the model. The model performs object detection on the input images and outputs the category label and bounding box for each category.
[0068] Record the detection results of the model, which include the detected target category, bounding box information, and confidence score;
[0069] After the model is trained, evaluate its performance on a separate test set.
[0070] The test set includes multiple categories of incremental scenarios to ensure that the model can exhibit good anti-forgetting ability under different incremental rhythms.
[0071] The calculated mean accuracy (mAP) was used to evaluate the model's detection performance on both new and old categories and its cross-task generalization ability, and to verify the model's effectiveness in mitigating catastrophic forgetting.
[0072] Compared with the prior art, the beneficial effects of the present invention are:
[0073] 1. This invention alleviates catastrophic forgetting and stably retains the detection performance of old categories when learning new categories without reusing old task data;
[0074] 2. This invention eliminates knowledge bias by learning cross-task relationship consistency, reducing the model's dependence on task-specific semantics and adapting to the semantic differences of different tasks.
[0075] 3. This invention stabilizes the feature structure and maintains clear category boundaries through feature distribution constraints, preventing new category features from destroying the feature space of old tasks;
[0076] 4. This invention improves cross-task adaptability, strengthens intra-class and inter-class relationship learning, and enhances the model's generalization ability in single-step and multi-step incremental scenarios.
[0077] 5. This invention simplifies the training process, designs a collaborative optimization framework, balances training efficiency and performance, and promotes the practical application of incremental detection technology. Attached Figure Description
[0078] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0079] Figure 1 This is a system architecture diagram of the present invention. Detailed Implementation
[0080] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0081] This invention aims to solve the catastrophic forgetting problem caused by knowledge shift in existing incremental object detection methods. It provides a method that balances cross-task shared semantic learning with the protection of feature structure integrity, so that the model can stably retain the detection performance of old categories when learning new categories of objects, thereby improving the robustness and generalization ability of incremental learning.
[0082] Please see Figure 1 A method for mitigating catastrophic forgetting in incremental target detection based on incremental relationship consistency includes at least the following steps:
[0083] S1: Perform data processing and initialization;
[0084] S2: Construct a Siamese Contrast Learning (SCRL) module. The query image and support images are input into a weight-shared ResNet-50 backbone network, which outputs the basic features of the query. and supporting basic features Where c is the number of categories and d is the feature dimension; for A multi-head self-attention mechanism is applied to enhance the pixel correlation within the feature map and suppress background noise; the supporting features are downsampled to a fixed size of 7×7 through pooling operations to unify the dimensions and reduce the computational complexity of subsequent operations.
[0085] S3: Build a boundary-aware structure preservation module, also known as the BASP module. The BASP module maintains the class center and decision boundary by constraining the loss function. This helps to reserve space for future tasks and explicitly restrict the feature distribution range of the old class during incremental training.
[0086] S4: Use total loss to optimize the overall objective;
[0087] S5: Conduct model evaluation and testing.
[0088] S1 includes at least the following steps:
[0089] Before each incremental task is started, for the t-th incremental task, the ABR (AugmentedBoxReplay) method is used to select representative target regions from the new task data to crop images and form a candidate set of supporting images. Then, it is merged with the historical task cache to obtain a complete support set B covering all old and new categories.
[0090] Next, the model parameters are initialized:
[0091] The first task initializes the ResNet-50 backbone network parameters θ and the Faster R-CNN classifier / regressor head parameters φ.
[0092] When it is a subsequent incremental task, it inherits the parameters θt-1 and φt-1 after training the previous task, and initializes the parameters required by the BASP module, such as class center and class boundary radius r.
[0093] Finally, the input data is divided, and the complete image to be detected in the current task is used as the query image (for final detection), while the target cropped image randomly selected from the memory cache B is used as the support image (to provide cross-task feature reference).
[0094] The application of the SCRL module includes at least the following steps:
[0095] A randomly initialized task code T is introduced, and the features of the query set interact with the new task code to learn new category specificities while preserving general relationships;
[0096] Cross-attention calculation is represented as:
[0097]
[0098] Where W1, W2, and W3 are learnable attention parameters, and A T Support features after interaction (only general relational semantics are retained). For query matrix; The key matrix;
[0099] Output features are processed using a single-head attention mechanism. The similarity between the query features and supporting features is calculated, and features irrelevant to the supported categories are filtered out. The matching coefficient is calculated as follows:
[0100]
[0101] Where W is The linear projection shared with Q, This represents the activation function. Represents element-level multiplication. Represents the query matrix;
[0102] The attention module is used to accurately identify the semantic relevance between query features and supporting features in object detection.
[0103] The fused features are then used for detection, following an inclusive distillation loss, and distillation is performed on the output to enhance the category knowledge of the old task:
[0104]
[0105] in, It is a normalization factor;
[0106] For the old task category The distillation loss at the current time step represents the current time step. The target category and related categories (from) -1 to The relationship between the probabilities of )
[0107] For all categories The loss calculation involves comparing the class information from the previous time step with the class probability of the current time step.
[0108] By comparing the prediction probabilities of the old model for supporting features with the prediction probabilities of the new model for fused features, the new model is forced to retain the old task category relationships to avoid knowledge forgetting.
[0109] The application of the BASP module includes at least the following steps:
[0110] The first task training loss function is proposed in the classification head of the detector, denoted as:
[0111]
[0112] in, An index representing all categories; Representative category The sample set; Indicates the size of the set; Indicates sample eigenvectors; It is a category The central eigenvector; Represents a constant; This represents the temperature parameter, used to control the smoothness of the loss function;
[0113] This allows new task features to be integrated into the boundary-defined feature space without interfering with the previously learned class distribution. This loss ensures that the feature space structure remains consistent throughout the incremental learning process.
[0114] After completing the training for the first task, estimate the boundary of the feature distribution for each category to determine the reserved boundary for subsequent tasks;
[0115] To simplify the analysis, the feature distribution of each category is modeled as follows: , The mean, For variance;
[0116] The boundary should be defined to include 90% of the distribution within the boundary; the corresponding hyperparameter is... (Right now ),and The time-based model has the best detection performance;
[0117] The boundary can be calculated as follows:
[0118]
[0119]
[0120] in It's a hyperparameter. The inverse cumulative distribution function of the standard normal distribution;
[0121] Used to calculate the upper and lower boundaries of the category feature distribution;
[0122] Through linear transformation, it is possible to determine Distribution median probability The corresponding left and right coordinate values;
[0123] Calculate the center threshold of the confidence interval based on the margin:
[0124]
[0125] The center threshold defines the decision boundary of the old task in the feature space, preserves the core region of the old task's feature space, and creates a buffer for subsequent tasks.
[0126] Calculate all boundary thresholds Then, the maximum historical boundary is defined as Used to dynamically adjust the marginal parameters in the loss function:
[0127]
[0128] Constraints Ensure there is a boundary between the center of the new task category and the maximum boundary of the old task. The minimum safe distance, this constraint ensures that there is a minimum safe distance between the center of the new task category and the maximum boundary of the old task. The minimum safe distance is used to prevent new category features from interfering with the distribution of old task features.
[0129] The total loss includes regression loss, classification loss from BASP, and inclusive distillation loss. .
[0130] S5 includes at least the following steps:
[0131] The test set images of the Pascal VOC and MS COCO benchmark datasets are input into the model. The model performs object detection on the input images and outputs the category label and bounding box for each category.
[0132] Record the model's detection results, which include the detected target category, bounding box information, and confidence score;
[0133] After the model is trained, evaluate its performance on a separate test set.
[0134] The test set includes multiple categories of incremental scenarios to ensure that the model can demonstrate good anti-forgetting ability under different incremental rhythms;
[0135] The calculated mean accuracy (mAP) was used to evaluate the model's detection performance on both new and old categories and its cross-task generalization ability, and to verify the model's effectiveness in mitigating catastrophic forgetting.
[0136] In summary:
[0137] This invention achieves the dual goals of cross-task relationship consistency learning and feature structure protection through the synergistic effect of SCRL and BASP, significantly mitigating catastrophic forgetting in incremental detection.
[0138] The SCRL module uses bi-branch contrastive learning and cross-attention fusion to shift the model from relying on task-specific semantics to learning general relational topology, effectively eliminating knowledge bias and improving cross-task adaptability. The BASP module uses dynamic boundary constraints to stabilize the distribution of old category features while reserving space for new category features, avoiding feature space distortion and enhancing model stability.
[0139] (1) Siamese contrastive relation learning (SCRL): Through a dual-branch structure and cross-attention fusion, it extracts intra-class and inter-class relation features that are universal across tasks, reducing task-specific semantic dependence. The core is to shift feature learning from "task-specific" to "relationship-universal", thereby alleviating knowledge bias from the root.
[0140] (2) Boundary-aware structure preservation (BASP): Based on the feature distribution constraints of dynamic margin parameters, the loss function is adjusted by calculating the boundary threshold of old category features to achieve a balance between "preservation of old feature structure" and "adaptation of new feature embedding" and avoid ambiguity of category boundaries.
[0141] (3) Adaptive memory buffering mechanism: A support set is constructed through a representative sample selection strategy. Without reusing old data, high-quality cross-task support samples are provided for the SCRL module, taking into account both memory efficiency and learning effect.
[0142] (4) Collaborative loss optimization framework: integrate detection loss, SCRL relation distillation loss and BASP boundary constraint loss, and balance the new task learning and old knowledge retention by balancing the loss terms through dynamic weight coefficients, so as to ensure the stability of the model in single-step and multi-step incremental scenarios.
[0143] These protective points work together to address the core shortcomings of existing incremental detection models. The key innovation lies in the dual design of "relational consistency learning + structural stability constraints," which breaks through the limitations of traditional distillation methods that rely on feature alignment and achieves efficient mitigation of catastrophic forgetting.
[0144] Experimental results confirm the superiority of this invention:
[0145] In the 10-10 single-step incremental setting of PascalVOC, the overall mAP reached 75.6%, and the mAP for new categories reached 76.7%, which is 3.5 percentage points higher than the existing best method GMDP-ABR (mAP for new categories = 73.2%). In the 10-1 multi-step incremental setting, the mAP is improved by 5 percentage points compared with competitors, and the performance degradation of old categories is reduced by more than 40%. In the 40-40 setting of MSCOCO, the AP reached 37.7%, which is better than advanced methods such as GMDP-ABR. Especially in complex weather and nighttime scenarios, the detection accuracy of core categories such as cars and pedestrians is significantly improved, which fully verifies the effectiveness of the present invention in practical applications.
[0146] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A method for mitigating catastrophic forgetting in incremental target detection based on incremental relationship consistency, characterized by: At least the following steps are included: S1: Perform data processing and initialization; S2: Construct a Siamese contrastive relation learning module, namely the SCRL module. The query image and support images are respectively input into a weight-shared ResNet-50 backbone network, and the output is the basic features of the query. and supporting basic features Where c is the number of categories and d is the feature dimension; for A multi-head self-attention mechanism is applied to enhance the pixel correlation within the feature map and suppress background noise; the supporting features are downsampled to a fixed size of 7×7 through pooling operations to unify the dimensions and reduce the computational complexity of subsequent operations. S3: Construct a boundary-aware structure preservation module, namely the BASP module, which maintains the class center and decision boundary through a constraint loss function; S4: Use total loss to optimize the overall objective; S5: Conduct model evaluation and testing.
2. The method for mitigating catastrophic forgetting in incremental target detection based on incremental relationship consistency according to claim 1, characterized in that: S1 includes at least the following steps: Before each incremental task is started, for the t-th incremental task, the ABR method is used to select representative target regions from the new task data to crop images and form a candidate set of supporting images. Then, it is merged with the historical task cache to obtain a complete support set B covering all old and new categories. Next, the model parameters are initialized: The first task initializes the ResNet-50 backbone network parameters θ and the Faster R-CNN classifier / regressor head parameters φ. When it is a subsequent incremental task, it inherits the parameters θt-1 and φt-1 after training the previous task, and initializes the parameters required by the BASP module, such as class center and class boundary radius r. Finally, the input data is divided, and the complete image to be detected in the current task is used as the query image, while the target cropped image randomly selected from the memory cache B is used as the support image.
3. The method for mitigating catastrophic forgetting in incremental target detection based on incremental relationship consistency according to claim 2, characterized in that: The application of the SCRL module includes at least the following steps: A randomly initialized task code T is introduced, and the features of the query set interact with the new task code to learn new category specificities while preserving general relationships; Cross-attention calculation is represented as: Where W1, W2, and W3 are learnable attention parameters, and A T For the support features after interaction, For query matrix; The key matrix; Output features are processed using a single-head attention mechanism. The similarity between the query features and supporting features is calculated, and features irrelevant to the supported categories are filtered out. The matching coefficient is calculated as follows: Where W is The linear projection shared with Q, This represents the activation function. Represents element-level multiplication. Represents the query matrix; The attention module is used to accurately identify the semantic relevance between query features and supporting features in object detection. The fused features are then used for detection, following an inclusive distillation loss, and distillation is performed on the output to enhance the category knowledge of the old task: in, It is a normalization factor; For the old task category The distillation loss at the current time step represents the current time step. The target category and related categories (from) -1 to The relationship between the probabilities of ) For all categories The loss calculation involves comparing the class information from the previous time step with the class probability of the current time step.
4. The method for mitigating catastrophic forgetting in incremental target detection based on incremental relationship consistency according to claim 3, characterized in that: The application of the BASP module includes at least the following steps: The first task training loss function is proposed in the classification head of the detector, denoted as: in, An index representing all categories; Representative category The sample set; Indicates the size of the set; Indicates sample eigenvectors; It is a category The central eigenvector; Represents a constant; This represents the temperature parameter, used to control the smoothness of the loss function; After completing the training for the first task, estimate the boundary of the feature distribution for each category to determine the reserved boundary for subsequent tasks; To simplify the analysis, the feature distribution of each category is modeled as follows: , The mean, For variance; The boundary should be defined to include 90% of the distribution within the boundary; the corresponding hyperparameter is... ,and The time-based model has the best detection performance; The boundary can be calculated as follows: in It's a hyperparameter. The inverse cumulative distribution function of the standard normal distribution; Used to calculate the upper and lower boundaries of the category feature distribution; Through linear transformation, it is possible to determine Distribution median probability The corresponding left and right coordinate values; Calculate the center threshold of the confidence interval based on the marginal values: The center threshold defines the decision boundary of the old task in the feature space, preserves the core region of the old task's feature space, and creates a buffer for subsequent tasks. Calculate all boundary thresholds Then, the maximum historical boundary is defined as Used to dynamically adjust the marginal parameters in the loss function: Constraints Ensure there is a boundary between the center of the new task category and the maximum boundary of the old task. The minimum safe distance, this constraint ensures that there is a minimum safe distance between the center of the new task category and the maximum boundary of the old task. The minimum safe distance is used to prevent new category features from interfering with the distribution of old task features.
5. The method for mitigating catastrophic forgetting in incremental target detection based on incremental relationship consistency according to claim 4, characterized in that: The total loss includes regression loss, classification loss from BASP, and inclusive distillation loss. .
6. The method for mitigating catastrophic forgetting in incremental target detection based on incremental relationship consistency according to claim 5, characterized in that: The S5 includes at least the following steps: The test set images of the Pascal VOC and MS COCO benchmark datasets are input into the model. The model performs object detection on the input images and outputs the category label and bounding box for each category. Record the detection results of the model, which include the detected target category, bounding box information, and confidence score; After the model is trained, evaluate its performance on a separate test set. The test set includes multiple categories of incremental scenarios to ensure that the model can exhibit good anti-forgetting ability under different incremental rhythms. The calculated mean accuracy metric was used to evaluate the model's detection performance on both new and old categories and its cross-task generalization ability, and to verify the model's effectiveness in mitigating catastrophic forgetting.