A Dynamic Equilibrium Annotation Method and System for Large Visual Models

Abstract

This invention discloses a dynamic balanced annotation method and system for large-scale visual models. The method includes: S1: First, calculate the label gap value, then determine if the label gap value is greater than zero. If it is, calculate the priority weight and push it to the high-priority queue, and label the database; otherwise, mark it as a saturated category and execute step S2; S2: First, generate multiple subsets, then train multiple model instances in parallel, subsequently plot the AP-sample size curve, and finally, when the AP growth rate of multiple consecutive samples is less than a preset percentage, determine that the inflection point sample size is the saturation value; S3: First, determine if the number of available labels exceeds the saturation value. If it is, randomly sample samples for training; otherwise, use all samples for training, and then feed the training results back to step S1 to adjust the indicators. This invention achieves precise matching between annotation resources and model requirements, significantly reduces computational load, greatly improves the ability to recognize small sample categories, and solves the problems of static resource allocation waste and blind spots in sample utility.

Description

Technical Field

[0001] This invention relates to a dynamic equilibrium annotation method in the field of model annotation technology, particularly to a dynamic equilibrium annotation method for large visual models, and also to a dynamic equilibrium annotation system for large visual models. Background Technology

[0002] In the field of large-scale visual model annotation, traditional static allocation mechanisms suffer from severe resource misallocation. A typical scenario is power grid equipment inspection, where fixed task queues cannot respond to the real-time demands of model training: while the detection rate for the "insulator crack" category is only 52% due to insufficient samples, the "pollution deposition" category wastes 30% of annotation resources due to excessive samples. Furthermore, existing systems lack the ability to quantify sample utility, making it difficult to determine when to terminate annotation—a security dataset shows that when the sample size exceeds the utility threshold, the model's AP value increases by less than 0.5%, but traditional solutions continue to allocate 40% of annotation resources to saturated categories, leading to a sharp decline in marginal benefits.

[0003] Current mainstream technical solutions have fundamental limitations: (1) Static process defects (such as patent CN120196947A), which adopts a fixed process of "large model pre-annotation + manual verification" and cannot dynamically schedule resources based on model feedback; (2) Lack of utility perception (such as patent CN116721419A), which relies on manual interaction to confirm target blocks and has not established a quantitative constraint mechanism for diminishing marginal utility of samples; (3) Insufficient scenario adaptation (such as patent CN119443684B), whose dynamic scheduling method focuses on the field of project management and does not combine the real-time performance feedback requirements in visual model training. These defects lead the annotation system into a vicious cycle of "resource mismatch → invalid annotation → model performance bottleneck". Summary of the Invention

[0004] To address the technical problems of static resource allocation waste and blind spots in sample utility in existing annotation methods, this invention provides a dynamic equilibrium annotation method and system for large visual models.

[0005] This invention employs the following technical solution: a dynamic equilibrium annotation method for large visual models, comprising the following steps:

[0006] S1: First, input the start value, the number of available tags, and the redundancy coefficient. Calculate the tag gap value. Then, determine whether the tag gap value is greater than zero. If it is, calculate the priority weight and push it to the high-priority queue and label the database. Otherwise, mark it as a saturated category and execute step S2. Wherein, the tag gap value = the start value × (1 + the redundancy coefficient) - the number of available tags.

[0007] S2: First, generate multiple subsets with successively decreasing sample sizes, then train multiple model instances corresponding to the multiple subsets in parallel, evaluate the AP index of each model, then plot the AP-sample size curve, and finally determine the corresponding inflection point sample size as the saturation value when the AP growth rate of multiple consecutive samples is less than a preset percentage.

[0008] S3: First determine whether the number of available labels exceeds the saturation value. If yes, randomly select a number of samples equal to the saturation value from all samples for training. Otherwise, use all samples for training and then feed the training results back to step S1 to dynamically adjust the starting value and the redundancy coefficient.

[0009] This invention solves the resource mismatch problem of static scheduling by linking label gap calculation with resource allocation in real time, filters invalid samples through physical blocking, quantifies the sample utility boundary, and forms a closed-loop evolution system through steps S1-S3. The evaluation index feedback drives parameter self-optimization, adapts to model architecture upgrades and scenario changes, and solves the technical problems of static resource allocation waste and sample utility blind spots in existing labeling methods.

[0010] As a further improvement to the above scheme, in step S1, the maximum value of the label gap value of all categories is first determined, and then each label gap value of the positive category is divided by the maximum value for normalization processing to obtain the corresponding priority weights. Finally, the categories are sorted in descending order of the priority weights to generate the priority queue.

[0011] As a further improvement to the above scheme, in step S1, visual monitoring is also performed: the label gap value status of each category is displayed in real time through a heat map; wherein, when the label gap value is greater than a preset high gap value, the heat map is a red heat map; when the label gap value is between the preset high gap value and a preset medium gap value, the heat map is a yellow heat map; when the label gap value is between the preset medium gap value and 0, the heat map is a green heat map, allowing labeling personnel to intuitively grasp the dynamics of resource allocation.

[0012] Furthermore, in step S3, the AP index and detection rate obtained from training are fed back to step S1 in real time, and the activation value is adjusted according to a preset dynamic adjustment rule.

[0013] Furthermore, when the heatmap is green and the detection rate is lower than a preset ratio, the redundancy coefficient is increased to a preset coefficient; when the heatmap is red and the label gap values ​​in multiple rounds are all greater than the preset high gap value, an emergency labeling task is automatically triggered.

[0014] Furthermore, the initial value of the redundancy coefficient is 0.1, the preset high gap value is 500, the preset medium gap value is 100, the preset ratio is 90%, and the preset coefficient is 0.15.

[0015] As a further improvement to the above scheme, the priority queue generation process includes a hardware acceleration mechanism, which uses an FPGA chip to calculate the label gap values ​​of all categories in parallel and completes the normalized weight sorting within a preset time.

[0016] Furthermore, the preset dynamic adjustment rules specifically include: (1) if the AP index of a certain category is ≥95% for 3 consecutive training cycles, the corresponding starting value will be increased by 20%; (2) if the label gap value of a certain category is continuously negative and the AP index fluctuates <0.3%, the corresponding starting value will be automatically reduced by 10%.

[0017] Furthermore, the execution method of the emergency annotation task is as follows: (1) calling the pre-trained large model to generate candidate annotation regions; (2) adopting a two-person cross-validation mechanism, when the annotation results of the two people are consistent, they are directly entered into the database, and when there is a disagreement, the arbitration module makes the determination; (3) after the task is completed, the extra computing resources occupied are released and the redundancy coefficient is restored to the initial value.

[0018] This invention also provides a dynamic equilibrium annotation system for large visual models, which applies any of the above-described dynamic equilibrium annotation methods for large visual models; the system includes:

[0019] Tag notch value driving module;

[0020] Saturation value execution module;

[0021] Closed-loop feedback components;

[0022] The label gap value driving module is used to first input the starting value, the number of available labels, and the redundancy coefficient, calculate the label gap value, and then determine whether the label gap value is greater than zero. If it is, the priority weight is calculated and pushed to the high-priority queue and labeled in the database; otherwise, it is marked as a saturated category and the saturation value execution module is executed. The label gap value = the starting value × (1 + the redundancy coefficient) - the number of available labels. The saturation value execution module is used to first generate multiple subsets with successively decreasing sample sizes, then train multiple model instances corresponding to the multiple subsets in parallel, evaluate the AP index of each model, then draw the AP-sample size curve, and finally determine the corresponding inflection point sample size as the saturation value when the AP growth rate of multiple consecutive samples is less than a preset percentage. The closed-loop feedback component is used to first determine whether the number of available labels exceeds the saturation value. If it is, it randomly selects samples equal to the number of saturation values ​​from all samples for training; otherwise, it uses all samples for training and then feeds back the training results to the label gap value driving module to dynamically adjust the starting value and the redundancy coefficient.

[0023] Compared with existing annotation methods, the dynamic equilibrium annotation method and system for large visual models of the present invention have the following beneficial effects:

[0024] 1. This dynamic equilibrium annotation method for large visual models solves the resource mismatch problem of static scheduling by real-time linkage between label gap calculation and resource allocation. It filters invalid samples through physical blocking, quantifies the sample utility boundary, and forms a closed-loop evolution system through steps S1-S3. The evaluation index feedback drives parameter self-optimization, adapts to model architecture upgrades and scene changes, and solves the technical problems of static resource allocation waste and sample utility blind spots in existing annotation methods.

[0025] 2. This dynamic equilibrium annotation method for large visual models achieves precise matching between annotation resources and model requirements through dynamic priority allocation driven by label gap values. Examples show that in power grid defect detection scenarios, resource utilization is increased from 58% to 92% compared to traditional methods, and the annotation speed for key categories (such as <1mm cracks) is accelerated by 40%, completely resolving the resource mismatch problem caused by human experience.

[0026] 3. This dynamic equalization annotation method for large visual models physically blocks invalid samples from entering the training process through a hard constraint mechanism of saturation values, significantly reducing the computational load. Examples demonstrate that GPU training time is reduced from 4.1 hours to 1.5 hours (a 63% reduction), and the model iteration cycle is compressed to one-third of the original under the same computing power conditions.

[0027] 4. This dynamic equilibrium labeling method for large-scale visual models prioritizes the supply of scarce sample resources through its label gap value mechanism, significantly improving the ability to identify small sample categories. In power grid inspection scenarios, the detection rate of low-sample categories (insulator cracks) jumped from 51% to 89%, while the overall false detection rate of the model was stably controlled below 3%.

[0028] 5. This dynamic equalization annotation method for large visual models uses evaluation metrics to drive dynamic parameter adjustments (e.g., starting value ±20%, redundancy coefficient fluctuating from 0.05 to 0.15), enabling the system to continuously optimize. When the model architecture is upgraded (e.g., ViT replaces ResNet) or the scene changes (power grid → security), parameter self-adaptation can be completed within 3 training cycles without manual intervention.

[0029] 6. This dynamic equilibrium annotation method for large-scale visual models reduces the number of invalid labeled samples by 37% (750 vs 1200 samples in the example), and reduces the manual review time from 40 hours / 10,000 images to 12 hours / 10,000 images. At a sample scale of millions, the overall annotation cost is reduced by 57% (calculated based on a 68% reduction in manual annotation costs + a 46% reduction in GPU energy consumption costs).

[0030] 7. The beneficial effects of this dynamic equilibrium annotation system for large visual models are the same as those of the dynamic equilibrium annotation method described above, and will not be repeated here. Attached Figure Description

[0031] Figure 1 This is a system architecture flowchart of the dynamic equalization annotation method for large visual models according to Embodiment 1 of the present invention.

[0032] Figure 2 for Figure 1 The flowchart of Δ value calculation for the dynamic equilibrium annotation method for large visual models is shown in the figure.

[0033] Figure 3 for Figure 1 The flowchart of the saturation value detection process for the dynamic equilibrium annotation method for large visual models is shown in the figure. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0035] Example 1

[0036] Please see Figure 1 , Figure 2 as well as Figure 3This embodiment provides a dynamic equilibrium annotation method for large-scale visual models. This annotation method is suitable for training large-scale visual models in complex scenarios such as power grid equipment detection and security monitoring, and establishes an intelligent closed loop of "gap calculation → resource allocation → effect verification". In this embodiment, the dynamic equilibrium annotation method includes the following steps (S1-S3).

[0037] S1: First, input the starting value, the number of available tags, and the redundancy coefficient. Calculate the tag gap value, then determine if the tag gap value is greater than zero. If it is, calculate the priority weight and push it to the high-priority queue, and label it in the database. Otherwise, mark it as a saturated category and execute step S2. Wherein, tag gap value = starting value × (1 + redundancy coefficient) - number of available tags. That is: Δ = c0 × (1 + α) - c1, where Δ represents the tag gap value, c0 represents the starting value, the redundancy coefficient represents α, and c1 represents the number of available tags. It should be noted that the redundancy coefficient in this embodiment has an initial value of 0.1 and can be dynamically adjusted according to scenario requirements. The priority sequence of labeling tasks is automatically adjusted according to Δ, with higher Δ categories receiving priority resource allocation.

[0038] In this embodiment, the maximum value among all category label gap values ​​is first determined. Then, each label gap value for a positive category is divided by the maximum value for normalization, resulting in corresponding priority weights, i.e., priority weights between 0 and 1. Finally, the categories are sorted in descending order of priority weights to generate a priority queue. For example, when the crack category label gap value is 800 and the contamination category label gap value is 180, the crack priority is 1.0 (800 / 800), and the contamination priority is 0.225 (180 / 800), with resources allocated in an 8:2 ratio. The priority queue generation process includes a hardware acceleration mechanism, which can use an FPGA chip to calculate the label gap values ​​of all categories in parallel and complete the normalized weight sorting within a preset time.

[0039] In addition, step S1 also includes visual monitoring: a heatmap is used to display the label gap value status of each category in real time. Specifically, when the label gap value is greater than a preset high gap value, the heatmap is red. When the label gap value is between the preset high gap value and a preset medium gap value, the heatmap is yellow. When the label gap value is between the preset medium gap value and 0, the heatmap is green, allowing labelers to intuitively grasp the dynamics of resource allocation. As one implementation, the preset high gap value can be 500, and the preset medium gap value can be 100. That is, Δ is visualized through the heatmap, displaying the category gap status as Δ>500 (red), 100<Δ≤500 (yellow), and Δ≤0 (green).

[0040] When the heatmap is green and the detection rate is lower than a preset ratio, the redundancy coefficient is increased to a preset coefficient to expand the redundancy space for calculating the label gap value and prioritize resource allocation. When the heatmap is red and the label gap value in multiple rounds is greater than the preset high gap value, an emergency labeling task is automatically triggered. In this embodiment, the preset ratio is 90% and the preset coefficient is 0.15. The execution method of the emergency labeling task can be: (1) calling the pre-trained large model to generate candidate labeling regions; (2) adopting a two-person cross-validation mechanism, when the labeling results of the two people are consistent, they are directly entered into the database, and when there is a disagreement, the arbitration module makes the judgment; (3) after the task is completed, the extra computing resources occupied are released and the redundancy coefficient is restored to the initial value.

[0041] S2: First, generate multiple subsets with successively decreasing sample sizes, then train multiple model instances corresponding to the multiple subsets in parallel, evaluate the AP metric of each model, then plot the AP-sample size curve, and finally determine the corresponding inflection point sample size as the saturation value when the AP growth rate of multiple consecutive samples is less than a preset percentage.

[0042] In this embodiment, the saturation value is determined through comparative training experiments: First, four subsets with sample sizes of 100%, 80%, 60%, and 40% are generated, and the model is trained in parallel while calculating the AP index; then, an AP-sample size curve is plotted, and when the AP increase over 1000 consecutive samples is less than 0.5%, the sample size is determined to be the saturation value. For example, for the crack category, the AP value reaches 95.3% with a sample size of 850, and the subsequent sample increase is only 0.2%, so the saturation value is set to 850.

[0043] S3: First, determine if the number of available labels exceeds the saturation value. If so, randomly select samples equal to the saturation value from all samples for training; otherwise, use all samples for training. Then, feed the training results back to step S1 to dynamically adjust the start value and redundancy coefficient. For example, if the saturation value for the contamination category is 560, even if 700 labels are labeled, only 560 samples are selected for training. After the model training is complete, the evaluation module outputs indicators such as AP value and detection rate, which are fed back to the Δ value calculation engine in real time. The engine dynamically adjusts the start value based on the indicators (e.g., if the detection rate of a certain category exceeds 95% for three consecutive periods, the start value is automatically increased by 20%), and the updated Δ value is passed to the resource allocator to achieve dynamic priority updates.

[0044] In this embodiment, the AP index and detection rate obtained from training are fed back to step S1 in real time, and the starting value is adjusted according to a preset dynamic adjustment rule. The preset dynamic adjustment rule specifically includes: (1) if the AP index of a certain category is ≥95% for 3 consecutive training cycles, the corresponding starting value is increased by 20%; (2) if the label gap value of a certain category is continuously negative and the AP index fluctuates <0.3%, the corresponding starting value is automatically reduced by 10%.

[0045] In summary, compared with existing annotation methods, the dynamic equilibrium annotation method for large visual models in this embodiment has the following advantages:

[0046] This dynamic equilibrium annotation method for large visual models solves the resource mismatch problem of static scheduling by linking label gap calculation with resource allocation in real time. It filters invalid samples through physical blocking, quantifies the sample utility boundary, and forms a closed-loop evolution system through steps S1-S3. The evaluation index feedback drives parameter self-optimization, adapts to model architecture upgrades and scene changes, and solves the technical problems of static resource allocation waste and sample utility blind spots in existing annotation methods.

[0047] Example 2

[0048] This embodiment provides a dynamic equilibrium annotation method for large visual models, which performs power grid insulator defect detection based on Embodiment 1.

[0049] I. System Initialization

[0050] Configure target category parameters: insulator cracks (starting value 1000, tag gap value 0.1), contamination deposition (starting value 800, tag gap value 0.1), with initial saturation values ​​set to 0 for both. Set evaluation index thresholds (AP ≥ 0.95, detection rate ≥ 0.90).

[0051] II. Label Gap Value Calculation and Resource Allocation

[0052] Input available tag data: 300 cracks, 700 stains, calculate tag notch value (Δ):

[0053] Crack Δ = 1000 × (1 + 0.1) - 300 = 800 (Red heat map, high notch)

[0054] Dirt Δ=800×(1+0.1)-700=180 (yellow heat map, gap in the middle)

[0055] Priority ranking results: Crack (1.0) → Dirt (0.225), and the annotation resources are allocated in an 8:2 ratio, that is, 80% of the manpower is given priority to handle crack annotation.

[0056] III. Saturation Value Detection and Application

[0057] Comparative training experiment statistics table

[0058]

[0059]

[0060] Inflection point identification: The AP of cracks increased by 0.013 (>0.5%) with 850 samples, but there was no significant improvement thereafter, and the saturation value was set at 850; the AP of dirt increased by 0.02 (>0.5%) with 560 samples, and the saturation value was set at 560.

[0061] Resource Release: The crack marking task terminates, releasing 60% of resources to the filth category. At this time, the filth marking priority is increased to 0.8 (Δ=800×1.1-560=320).

[0062] IV. Closed-loop verification

[0063] Comparison table of optimized indicators

[0064]

[0065] Based on the above results, this dynamic equilibrium annotation method for large visual models has the following advantages compared to existing technologies:

[0066] 1. This dynamic equilibrium annotation method for large visual models achieves precise matching between annotation resources and model requirements through dynamic priority allocation driven by label gap values. Examples show that in power grid defect detection scenarios, resource utilization is increased from 58% to 92% compared to traditional methods, and the annotation speed for key categories (such as <1mm cracks) is accelerated by 40%, completely resolving the resource mismatch problem caused by human experience.

[0067] 2. This dynamic equalization annotation method for large visual models physically blocks invalid samples from entering the training process through a hard constraint mechanism of saturation values, significantly reducing the computational load. Examples demonstrate that GPU training time is reduced from 4.1 hours to 1.5 hours (a 63% reduction), and the model iteration cycle is compressed to one-third of the original under the same computing power conditions.

[0068] 3. This dynamic equilibrium labeling method for large visual models prioritizes the supply of scarce sample resources through its label gap value mechanism, significantly improving the ability to identify small sample categories. In power grid inspection scenarios, the detection rate of low sample categories (insulator cracks) jumped from 51% to 89%, while the overall false detection rate of the model was stably controlled below 3%.

[0069] 4. This dynamic equalization annotation method for large visual models uses evaluation metrics to drive dynamic parameter adjustments (e.g., starting value ±20%, redundancy coefficient fluctuating from 0.05 to 0.15), enabling the system to continuously optimize. When the model architecture is upgraded (e.g., ViT replaces ResNet) or the scene changes (power grid → security), parameter self-adaptation can be completed within 3 training cycles without manual intervention.

[0070] 5. This dynamic equalization annotation method for large visual models reduces the number of invalid labeled samples by 37% (750 vs 1200 samples in the example), and reduces the manual review time from 40 hours / 10,000 images to 12 hours / 10,000 images. At a sample scale of millions, the overall annotation cost is reduced by 57% (calculated based on a 68% reduction in manual annotation costs + a 46% reduction in GPU energy consumption costs).

[0071] Example 3

[0072] This embodiment provides a dynamic equilibrium annotation system for large visual models. The system applies the dynamic equilibrium annotation method for large visual models in embodiment 1 or 2, and includes a label gap value driving module, a saturation value execution module, and a closed-loop feedback component.

[0073] In this embodiment, the label gap value driving module first inputs the starting value, the number of available labels, and the redundancy coefficient, calculates the label gap value, and then determines whether the label gap value is greater than zero. If it is, it calculates the priority weight and pushes it to the high-priority queue and labels the database; otherwise, it marks it as a saturated category and executes the saturation value execution module. The label gap value is calculated as: Starting value × (1 + Redundancy coefficient) - Number of available labels. The saturation value execution module first generates multiple subsets with progressively decreasing sample sizes, then trains multiple model instances corresponding to these subsets in parallel, evaluates the AP (Average Per Second) of each model, plots the AP-sample size curve, and finally determines the corresponding inflection point sample size as the saturation value when the AP growth rate of multiple consecutive samples is less than a preset percentage. The closed-loop feedback component first determines whether the number of available labels exceeds the saturation value. If so, it randomly selects samples equal to the saturation value from all samples for training; otherwise, it uses all samples for training and feeds the training results back to the label gap value driving module to dynamically adjust the starting value and redundancy coefficient.

[0074] Example 4

[0075] This embodiment provides a computer terminal, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the dynamic equalization annotation method for large visual models described in Embodiment 1 or 2.

[0076] When applying the methods of Embodiment 1 or 2, they can be implemented in software form, such as by designing them as stand-alone programs and installing them on a computer terminal, which can be a computer, smartphone, control system, or other IoT devices. Alternatively, the methods of Embodiment 1 or 2 can be designed as embedded programs and installed on a computer terminal, such as on a microcontroller.

[0077] Example 5

[0078] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, it implements the steps of the dynamic equalization annotation method for large visual models according to Embodiment 1 or 2.

[0079] When applying the method of Embodiment 1 or 2, it can be applied in the form of software, such as by designing it as a program that can run independently on a computer-readable storage medium. The computer-readable storage medium can be a USB flash drive, designed as a USB security token, and the program is designed to start the entire method through an external trigger.

[0080] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic equilibrium annotation method for large visual models, characterized in that, It includes the following steps: S1: First, input the start value, the number of available tags, and the redundancy coefficient. Calculate the tag gap value. Then, determine whether the tag gap value is greater than zero. If it is, calculate the priority weight and push it to the high-priority queue and label the database. Otherwise, mark it as a saturated category and execute step S2. Wherein, the tag gap value = the start value × (1 + the redundancy coefficient) - the number of available tags. S2: First, generate multiple subsets with successively decreasing sample sizes, then train multiple model instances corresponding to the multiple subsets in parallel, evaluate the AP index of each model, then plot the AP-sample size curve, and finally determine the corresponding inflection point sample size as the saturation value when the AP growth rate of multiple consecutive samples is less than a preset percentage. S3: First determine whether the number of available labels exceeds the saturation value. If yes, randomly select a number of samples equal to the saturation value from all samples for training. Otherwise, use all samples for training and then feed the training results back to step S1 to dynamically adjust the starting value and the redundancy coefficient.

2. The dynamic equilibrium annotation method for large visual models as described in claim 1, characterized in that, In step S1, the maximum value of the label gap values ​​of all categories is first determined. Then, each label gap value of the positive category is divided by the maximum value for normalization to obtain the corresponding priority weights. Finally, the categories are sorted in descending order of the priority weights to generate the priority queue.

3. The dynamic equilibrium annotation method for large visual models as described in claim 1, characterized in that, In step S1, visual monitoring is also performed: the label gap value status of each category is displayed in real time through a heat map; wherein, when the label gap value is greater than a preset high gap value, the heat map is a red heat map; when the label gap value is between the preset high gap value and a preset medium gap value, the heat map is a yellow heat map; when the label gap value is between the preset medium gap value and 0, the heat map is a green heat map, allowing labeling personnel to intuitively grasp the dynamics of resource allocation.

4. The dynamic equalization annotation method for large visual models as described in claim 3, characterized in that, In step S3, the AP index and detection rate obtained from training are fed back to step S1 in real time, and the activation value is adjusted according to a preset dynamic adjustment rule.

5. The dynamic equalization annotation method for large visual models as described in claim 4, characterized in that, When the heatmap is green and the detection rate is lower than a preset ratio, the redundancy coefficient is increased to a preset coefficient; when the heatmap is red and the label gap values ​​in multiple rounds are greater than the preset high gap value, an emergency labeling task is automatically triggered.

6. The dynamic equalization annotation method for large visual models as described in claim 5, characterized in that, The initial value of the redundancy coefficient is 0.1, the preset high gap value is 500, the preset medium gap value is 100, the preset ratio is 90%, and the preset coefficient is 0.

15.

7. The dynamic equilibrium annotation method for large visual models as described in claim 1, characterized in that, The priority queue generation process includes a hardware acceleration mechanism, which uses an FPGA chip to calculate the label gap values ​​of all categories in parallel and completes the normalized weight sorting within a preset time.

8. The dynamic equalization annotation method for large visual models as described in claim 4, characterized in that, The preset dynamic adjustment rules specifically include: (1) If the AP index of a certain category is ≥95% for 3 consecutive training cycles, the corresponding starting value will be increased by 20%; (2) If the label gap value of a certain category is continuously negative and the AP index fluctuates <0.3%, the corresponding starting value will be automatically reduced by 10%.

9. The dynamic equalization annotation method for large visual models as described in claim 5, characterized in that, The execution method of the emergency annotation task is as follows: (1) Call the pre-trained large model to generate candidate annotation regions; (2) Adopt a two-person cross-validation mechanism. When the annotation results of the two people are consistent, they are directly stored in the database. When there is a disagreement, the arbitration module makes the determination; (3) After the task is completed, release the extra computing resources occupied and restore the redundancy coefficient to the initial value.

10. A dynamic equilibrium annotation system for large visual models, characterized in that, Its application is the dynamic equilibrium annotation method for large visual models as described in any one of claims 1-9; the system includes: Tag gap value driving module; Saturation value execution module; Closed-loop feedback components; The label gap value driving module is used to first input the starting value, the number of available labels, and the redundancy coefficient, calculate the label gap value, and then determine whether the label gap value is greater than zero. If it is, the priority weight is calculated and pushed to the high-priority queue and labeled in the database; otherwise, it is marked as a saturated category and the saturation value execution module is executed. The label gap value = the starting value × (1 + the redundancy coefficient) - the number of available labels. The saturation value execution module is used to first generate multiple subsets with successively decreasing sample sizes, then train multiple model instances corresponding to the multiple subsets in parallel, evaluate the AP index of each model, then draw the AP-sample size curve, and finally determine the corresponding inflection point sample size as the saturation value when the AP growth rate of multiple consecutive samples is less than a preset percentage. The closed-loop feedback component is used to first determine whether the number of available labels exceeds the saturation value. If it is, it randomly selects samples equal to the number of saturation values ​​from all samples for training; otherwise, it uses all samples for training and then feeds back the training results to the label gap value driving module to dynamically adjust the starting value and the redundancy coefficient.

Images