A dynamic mass feedback generative data augmentation method

Abstract

The application discloses a dynamic quality feedback generative data enhancement method, relates to the field of data enhancement, and solves the technical problems of lack of semantic information of existing data enhancement technology and uncontrollability of a traditional GAN enhancement direction; the application comprises the following steps: taking a random noise vector and a condition vector as common inputs, inputting the common inputs into a generator to obtain generated samples; a discriminator compares real samples in an original training data set to be enhanced with the generated samples; the generator and the discriminator are alternately optimized through a loss function until an optimization condition is met; quality evaluation is performed on the samples generated by the optimized generator to obtain high-quality enhanced data; the high-quality enhanced data are applied to a downstream model, performance of the downstream model on a verification set is monitored, if the performance is insufficient, an adjustment strategy is determined according to the performance, and then the generated samples are returned to be regenerated; and the application significantly improves the pertinence and effectiveness of the enhanced data.

Description

Technical Field

[0001] This invention relates to the field of data augmentation, and more specifically to a generative data augmentation method with dynamic quality feedback. Background Technology

[0002] In the training of deep learning models, sufficient and high-quality labeled data is crucial for achieving good model performance. However, in practical applications, acquiring large amounts of labeled data is often costly and time-consuming, leading to insufficient training samples, making the model prone to overfitting and poor generalization ability. Data augmentation is a common technique to address this problem, expanding the dataset by performing a series of transformations on the original training data, such as rotation, scaling, cropping, and adding noise.

[0003] Existing data augmentation techniques are mostly rule-based low-level image transformations. The new samples generated differ from the original samples at the pixel level, but they do not introduce new semantic information. This type of augmentation has limited diversity and is insufficient in improving the model's generalization ability in complex scenarios. For example, in medical image analysis, simple geometric transformations cannot simulate the morphological changes at different stages of disease; in industrial quality inspection, it is also difficult to generate new samples with subtle defects and textures.

[0004] Generative Adversarial Networks (GANs), as powerful generative models, can learn the distribution of real data and generate realistic new samples, providing new ideas for data augmentation. However, the generation process of traditional GAN ​​models is stochastic and uncontrollable, making it impossible to generate targeted, semantically meaningful data for supervised learning tasks with discrete class labels (such as augmenting image samples of patients with specific diseases). This limits the application effectiveness of GANs in data augmentation scenarios that require precise control over the attributes of generated content. While existing technologies such as Conditional Generative Adversarial Networks (cGANs) introduce conditional control, they still suffer from problems such as unstable generation quality, low semantic alignment accuracy, and a lack of mechanisms for collaborative optimization with downstream tasks. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a generative data augmentation method with dynamic quality feedback, which solves the technical problems of insufficient semantic information in existing data augmentation techniques and uncontrollable augmentation direction in traditional GANs.

[0006] A generative data augmentation method with dynamic quality feedback includes the following steps:

[0007] S1: Receive a data augmentation request from an external training system. The data augmentation request includes a target semantic condition vector and a quantity requirement. Determine the category sampling weight distribution based on the target semantic condition vector. Sample semantic labels from a predefined category set, encode them, and use them as condition vectors.

[0008] S2: The random noise vector and the conditional vector are used as common inputs to the generator to generate samples; the discriminator compares the real samples and generated samples in the original training dataset to be enhanced; the generator and discriminator are alternately optimized through the loss function until the optimization condition is met;

[0009] S3: Perform quality assessment on the samples generated by the generator after S2 optimization to obtain high-quality augmented data;

[0010] S4: Apply high-quality augmented data to the downstream model, monitor the performance of the downstream model on the validation set, and if the performance is insufficient, determine the adjustment strategy based on the performance, and then return to S1 to generate samples again.

[0011] Furthermore, S1 includes:

[0012] Construct the probability distribution for the initial class sampling;

[0013] The category is selected based on the probability distribution;

[0014] The sampled category labels are encoded into condition vectors;

[0015] The conditional vector is input into the generator to control the semantic category of the generated samples.

[0016] Further, S3 includes:

[0017] The authenticity score is obtained by evaluating whether the generated sample approximates the real data in terms of feature distribution; the novelty score is obtained by evaluating whether the generated sample maintains a reasonable difference from the existing training data; the authenticity score and novelty score are compared with the corresponding thresholds, and if both are greater than the corresponding thresholds, they are judged as high-quality augmented data.

[0018] Furthermore, the evaluation of whether the generated samples approximate the real data in terms of feature distribution includes:

[0019] Calculate category feature centers: For each semantic category k, randomly select from its real training samples. For each sample, features are extracted and the mean is calculated to obtain the feature center of that category.

[0020] Calculate distance: For generated samples with condition label g Extract its features Calculate its distance to the corresponding category center Euclidean distance:

[0021] Convert to fractions: Mapping Euclidean distance to a fidelity score in the interval [0,1] using a negative exponential function:

[0022] Furthermore, the evaluation described in S3 regarding whether the generated samples maintain a reasonable difference from the existing training data includes:

[0023] Building a dynamic feature library: Building and maintaining a feature library The feature library It contains feature vectors of all real samples and historically qualified augmented samples.

[0024] Find the nearest neighbor distance: Calculate the features of the generated samples To the feature library The minimum Euclidean distance among all eigenvectors in the set is denoted as . .

[0025] Convert to a score: Map the distance to a novelty score in the interval [0,1] using the Sigmoid function.

[0026] Furthermore, the corresponding threshold is obtained by statistical analysis of samples in the original training dataset.

[0027] Furthermore, the analytical performance includes:

[0028] Analyzing the change in model accuracy before and after data augmentation as an indicator of overall performance change ;

[0029] The significance threshold is dynamically determined through statistical tests based on the validation set size. ;

[0030] Calculate the recall change for each category k It identifies weak and declining categories and performs inter-category confusion analysis; the weak category refers to... and <The category of average recall * x; where x = average recall - standard deviation of recall; the declining category refers to... The categories; the inter-class confusion analysis, comparing the confusion matrices before and after enhancement, finds the category pair (i,j) with the largest increase in the absolute difference of the false positive rate, and denoted as the confused category pair.

[0031] Furthermore, the determination of the adjustment strategy based on performance includes:

[0032] If overall performance declines The strategy of zero is adopted to achieve dynamic regression control of the training data distribution by jointly adjusting the strength of the generated distribution contraction, the sample selection threshold, the latent space variance, and the proportion of generated data, so that the augmented data distribution can approximate the real data distribution again, thereby restoring the model performance.

[0033] If the overall performance does not degrade but there are weak / degraded / confused class pairs, then strategy one, strategy two, and strategy three are adopted according to the existence of each class. Strategy one adjusts the conditional input of S1, strategy two adjusts the content of the original training dataset and the generation target of S2, and strategy three adjusts the threshold in the quality evaluation process of S3.

[0034] Furthermore, the strategy zero includes:

[0035] By adjusting the fidelity constraint parameters of the generated samples This makes the generated samples closer to the real data distribution;

[0036] Constrain the generator input noise distribution;

[0037] Control the proportion of augmented data samples to the total number of samples in each training batch of the downstream model.

[0038] Furthermore, Strategy 1 includes: for weak / declining categories, increasing their generation priority; for confused category pairs, generating boundary samples to refine the decision boundary; Strategy 2 includes: for weak categories, extracting features from samples misjudged by high loss in the weak category; generating feature-constrained loss based on these features and adding them to the loss function optimized by the regenerator; for confused category pairs, generating adversarial samples from the real samples in them.

[0039] The beneficial effects of this invention include:

[0040] This invention achieves semantically controllable data generation by introducing a Conditional Generative Adversarial Network (cGAN) architecture, overcoming the shortcomings of insufficient semantic information and random generation direction in traditional augmentation methods. A discriminator design combining realism and semantic discrimination ensures high-quality and semantically accurate generated samples. Furthermore, a quality assessment and filtering module and a dynamic augmentation feedback module form a closed-loop optimization system, enabling dynamic and adaptive data augmentation based on the training state of the downstream model. This significantly improves the relevance and effectiveness of the augmented data, thereby more efficiently enhancing the generalization performance and robustness of downstream deep learning models, especially in scenarios with scarce or imbalanced data. Attached Figure Description

[0041] Figure 1 This is a flowchart of a generative data augmentation method with dynamic quality feedback according to an embodiment of this application. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0043] Example 1

[0044] The following is in conjunction with the appendix Figure 1 Specific embodiments of the present invention will be described in detail;

[0045] A generative data augmentation method with dynamic quality feedback, comprising:

[0046] S1: Receive a data augmentation request from an external training system, the data augmentation request including a target semantic condition vector and a quantity requirement; determine the category sampling weight distribution based on the target semantic condition vector, increasing the weight of categories mentioned in the target semantic condition vector and decreasing the weight of those not mentioned; sample semantic labels from a predefined category set, encode them, and use them as condition vectors; including: assuming the category set is {k=1,2,...,K}, and the number of training samples corresponding to the Kth category is... The predefined category set is derived directly from the downstream classification task to be enhanced. During initialization, the original training dataset is parsed to obtain the labels of all samples in the dataset. The set of unique values ​​for these labels constitutes the predefined category set. The total number of samples is:

[0047]

[0048] (1) Construct the initial category sampling probability distribution:

[0049]

[0050] Where P(k) represents the probability that category k is sampled;

[0051] (2) Perform category sampling based on the probability distribution:

[0052] k~Categorical(P(k))

[0053] The sampling is achieved by constructing a cumulative distribution function and combining it with uniform random variables;

[0054] (3) The sampled category labels are encoded into conditional vectors using an embedding matrix mapping method:

[0055]

[0056] Where E(·) represents the encoding function, which maps discrete category labels to high-dimensional conditional vectors;

[0057] (4) The condition vector The input generator controls the semantic category of the generated samples.

[0058] S2: A random noise vector and a conditional vector are used as common inputs to the generator to produce generated samples. The discriminator compares the real samples with the generated samples, and alternately optimizes the generator and discriminator using a loss function. The generator is optimized to produce images that are both realistic and conform to the conditional labels. The discriminator is optimized to more accurately distinguish between real and generated images and determine the degree of matching between the image and the label. The optimization process terminates when any of the following conditions are met:

[0059] 1) Reach the preset maximum number of training rounds;

[0060] 2) The loss function changes below the preset threshold ε for several consecutive rounds.

[0061] The real samples are labeled and derived from the original training dataset to be augmented.

[0062] S3: Perform quality assessment on the samples generated by the generator after S2 optimization to obtain high-quality augmented data. The quality assessment includes: assessing whether the generated samples approximate the real data in terms of feature distribution; and assessing whether the generated samples maintain a reasonable difference from the existing training data.

[0063] S4: Apply high-quality augmented data to the downstream model, monitor the performance of the downstream model on the validation set, and adjust steps S1-S4 to generate samples again based on the performance.

[0064] In another embodiment, S1 includes: a user or downstream system sending an enhancement request specifying semantic conditions; a parser verifying the request format and extracting semantic conditions and quantity parameters; and an encoder converting text or tag-based conditions into high-dimensional vectors as condition inputs to the generator.

[0065] In another embodiment, S2 includes:

[0066] The generator generates samples by fusing random noise vectors and conditional vectors in the feature space to ensure that the generation direction is semantically controlled; gradually increasing the image size through transposed convolutional layers from 7×7 to 112×112, with each layer learning features at different scales; the conditional vectors participate in the calculation in each upsampling step to ensure that the output image is consistent with the target semantics; and generating 112×112 pixel RGB images for direct use in downstream tasks.

[0067] The discriminator compares real and generated samples, receiving two types of input: real training samples and their corresponding real labels, and synthetic samples generated by S2 and their corresponding conditional labels. The discriminator performs two tasks: a) determining whether the input image is real or generated (realism discrimination), outputting a probability value (0-1), where 0 represents a generated sample and 1 represents a real sample; b) determining whether the semantic category of the input image matches its conditional label (semantic discrimination), outputting a classification probability distribution to verify if the image matches the conditional label. The generator and discriminator are optimized using a loss function to improve the realism and semantic accuracy of the generated samples.

[0068] In another embodiment, S3 includes:

[0069] S31: For each generated sample to be evaluated Calculate the fidelity score and the novelty score;

[0070] 1. The fidelity score measures how closely the generated sample approximates the real data distribution. It is used to filter out off-target samples, such as images with distorted structures or unclear meanings. The calculation steps are as follows:

[0071] Calculate category feature centers: For each semantic category k, randomly select from its real training samples. For each sample, features are extracted and the mean is calculated to obtain the feature center of that category. :

[0072]

[0073] Calculate distance: For generated samples with condition label g Extract its features Calculate its distance to the corresponding category center Euclidean distance:

[0074]

[0075] Convert to a fraction: A fidelity score that maps the distance to the interval [0,1] using a negative exponential function.

[0076]

[0077] in, It is a scaling parameter used to adjust the sensitivity of the score to distance. Specifically, it adjusts the sensitivity of the score to changes in distance and normalizes the distance units for different datasets or feature spaces. The higher the value, the higher the realism score. With distance The faster it increases, the faster it decays.

[0078] It is the Euclidean distance from the generated sample features to the center of the true features of its corresponding class; The smaller the value, the closer the sample is to the core distribution of the real data in the feature space, and the higher its quality. When When =0, =1, indicating that the sample perfectly coincides with the true center, representing the highest fidelity; as... Increase Starting from 1 and monotonically decreasing towards 0, it indicates that the sample deviates more from the true distribution, and the fidelity decreases.

[0079] 2. The novelty score measures the novelty of the generated sample relative to the existing training set and is used to filter redundant samples. The existing training set includes the original data and historical augmented data that has passed filtering. The calculation steps are as follows:

[0080] Building a dynamic feature library: Building and maintaining a feature library The feature library It contains feature vectors of all real samples and augmented samples that have been filtered by this module.

[0081] Find the nearest neighbor distance: Calculate the features of the generated samples To the feature library The minimum Euclidean distance among all eigenvectors in the set is denoted as . .

[0082] Convert to a score: Novelty score by mapping the distance to the interval [0,1] using the Sigmoid function:

[0083]

[0084] in, It reflects the uniqueness of the sample; μ is the baseline distance, set as the median of the minimum distance between real samples in the feature library, used to distinguish between moderate differences and abnormal deviations; It is a parameter that adjusts the steepness of the Sigmoid function; when ( When ) is greater than 0, A value greater than 0.5 indicates that the sample is sufficiently novel; otherwise, it is too redundant.

[0085] and The value of is determined by calculating the distance from all real training samples to their class centers during the preprocessing stage. mean Minimum distance between and all real samples mean To make exp(- * )≈0.5 and sigmoid( * )≈0.75 (meaning the average distance corresponds to a score near the median), which can be set separately. =ln(2) / and =ln(3) / This setting ensures that the scores have clear discrimination at typical distances and are adaptable to different datasets.

[0086] S32: Define the lower limit threshold for realism Novelty lower limit threshold ; Used to reject samples that deviate from the target sample. If the sample is too low in quality, it will be filtered out directly. Used to reject redundant samples. If If the sample is not sufficiently informative, it will be filtered out. A sample is filtered only if it simultaneously satisfies... and When the sample is in good condition, it can be considered a high-quality sample.

[0087] First, the initial threshold is determined statistically based on the original training data:

[0088] Initial values: Calculate the values ​​of all real samples. The score is taken as the 5th percentile of its distribution. Correspondingly, the sample deviation refers to a lower fidelity than 95% of the real samples.

[0089] Initial values: Calculate the sum of all real samples and all other samples in the feature library except themselves. Take the 25th percentile of its distribution as μ, and set... for The neutral value. The sample redundancy refers to novelty being lower than the typical difference between real samples.

[0090] If S4 performance analysis requires adjustments to S3, dynamic fine-tuning is performed based on the downstream model's training status to achieve adaptive filtering. When S4 detects model overfitting, it provides feedback to improve... Tighten the fidelity requirements and reduce noisy samples; when S4 detects model underfitting or insufficient diversity, the feedback is appropriately reduced. This allows for a more diverse range of samples to be included; when S4 needs to focus on tackling a specific type of problem, it provides feedback to reduce the impact on that particular category. It allows difficult samples, such as those with severe occlusion or edge cases, to be included in the training to improve the robustness of the model.

[0091] In another embodiment, S4 includes:

[0092] S41: After each round of augmented data injection and retraining of the downstream model, initiate automated performance diagnostics on an independent validation set, analyzing the following dimensions:

[0093] 1. Overall performance change (ΔG):

[0094]

[0095] in This refers to the accuracy of the model after the previous round of enhancements, while the first round refers to the accuracy of the original model. ; This is the accuracy of the model after this round of enhancements.

[0096] 2. Set a significance threshold δ, which is dynamically determined through statistical tests based on the size of the validation set. Use the Z-test for proportional differences:

[0097]

[0098] Where n is the number of samples in the validation set, and z is the Z-value corresponding to the confidence level.

[0099] 3. Category Balance Analysis: Calculate the recall change for each category k. ,

[0100]

[0101] It identifies weak and declining categories and performs inter-class confusion analysis; the weak category refers to... and <The category of average recall * x; where x = average recall - standard deviation of recall; the declining category refers to... The categories; the inter-class confusion analysis, comparing the confusion matrices before and after enhancement, finds the category pair (i,j) with the largest increase in the absolute difference of the false positive rate, and denoted as the confused category pair.

[0102] S42: Strategic Decision-Making Logic;

[0103] Based on the S41 performance analysis results, the corresponding adjustment strategies are triggered in priority order:

[0104] P1. Performance degradation (ΔG < -δ); P2. Existence of weak class / degraded class / confused class pairs.

[0105] When overall performance degradation (ΔG < -δ) is detected, strategy zero is employed to dynamically regress the training data distribution by jointly adjusting the strength of the generated distribution contraction, the sample selection threshold, the latent space variance, and the proportion of generated data. This allows the augmented data distribution to approximate the true data distribution again, thereby restoring model performance. The steps include:

[0106] (1) Improve the quality constraints of generated samples

[0107] By adjusting the fidelity constraint parameters of the generated samples This makes the generated samples closer to the real data distribution:

[0108]

[0109] Where η>0 is the quality constraint enhancement coefficient, which is a dynamic adjustment value greater than 0. Its value should be positively correlated with the severity of performance degradation. In this method, it is set as follows:

[0110] η=0.1*∣ΔG∣ / δ

[0111] Since the corresponding fidelity score function is:

[0112]

[0113] By increasing , so that when the distance When they are a little older, This rapidly reduces the number of low-quality samples that deviate from the true distribution from the training set.

[0114] (2) Reduce the complexity of generation distribution

[0115] Constrain the input noise distribution of the cGAN generator:

[0116] In standard GANs / cGANs, the input to the generator G is typically a random noise vector z, sampled from a prior distribution, namely the standard normal distribution N(0,1). σ is a scaling factor for the variance of this distribution. That is, the actual sampled distribution is:

[0117] Z~N(0,σ 2 )

[0118] Adjust σ.

[0119] σ'=σ*(1-β)

[0120] Where β∈(0,1), in order to effectively reduce the extreme differences of generated samples and retain basic diversity, β=0.1 is set in this case. By reducing the noise variance σ, the distribution of generated samples is more concentrated, reducing training instability.

[0121] (3) Dynamically adjust the proportion of generated data

[0122] In each training batch of the downstream model, the proportion of samples from augmented data to the total number of samples is defined as ρ, and the proportion of augmented data ρ is controlled as follows:

[0123] ρ'=ρ×(1-γ)

[0124] Where γ∈(0,1) is the proportion decay coefficient. When performance degradation is first triggered, a large reduction is performed, such as setting γ=0.5 directly, i.e., ρ'=ρ / 2. In subsequent augmentation training loops, if performance begins to recover, i.e. ΔG>0, ρ is gradually restored to the original level with smaller steps. By reducing the proportion of augmented data, the model is trained with a more realistic data distribution, thereby improving accuracy.

[0125] When a weak category / declining category / confusion category pair is detected, strategies one, two, and three are implemented based on the presence of each category.

[0126] S43, Strategy 1: Adjust the conditional input of S1, including changing the sampling distribution weights of the conditional vector in S1, to guide the generation direction from the source:

[0127] For weak / declining categories: increase their generation priority. Let the current category sampling weight distribution be... For the category k that needs strengthening, its weights are updated:

[0128]

[0129] Where α>0 is the enhancement intensity factor.

[0130] Subsequently, all Normalize the weights so that their sum equals 1; the normalized weights will be used for conditional sampling during the next generation of the S1 module.

[0131] The enhancement strength factor α controls the intensity of enhancement for weaker categories. As an empirical, moderate adjustment value, it aims to avoid drastic changes in the training distribution and instability caused by excessively large single adjustments. In this embodiment, it is set to 0.3, which means that in each adjustment, the sampling weight of this category is increased to 1.3 times its original value.

[0132] For confused class pairs: generate boundary samples to refine the decision boundary; for key confused pairs (i,j), in addition to generating samples according to their original weights, the generator is further prompted to generate a sample with a conditional vector c, which lies between the two classes, challenging and strengthening the discriminative boundary.

[0133]

[0134] in Conditional encoding for categories i and j The mixing coefficient is the uniform sampling coefficient within the interval [0,1].

[0135] S44, Strategy 2: Adjust the content and generation target of the S2 training set to achieve more refined optimization by directly constructing or influencing the training data.

[0136] Feature constraint optimization: To address the weak class problem, this involves not only increasing the number of weak classes but also improving their quality. Features are extracted from samples that were misclassified with high loss by the model during historical training for that class. In the next round of training the S2 generator, the existing adversarial loss will be used... Based on this, a feature constraint loss is added for generated samples belonging to this category:

[0137] 2

[0138] Where E(·) is the feature extractor, This is the conditional vector for that category. The generator's total loss becomes:

[0139]

[0140] in, The weighting coefficients ensure that the feature constraint term serves as an important regularization guide, but does not dominate the generator's training, so as not to compromise the diversity of generated samples and the equilibrium of adversarial games. This loss term encourages the generator to produce samples that are close to historical hard examples in the feature space, thereby automatically learning to generate samples that are more challenging for the class and can more effectively compensate for model deficiencies.

[0141] Adversarial example enhancement: To alleviate class confusion and improve model robustness, adversarial examples are generated for the real sample x in the class confusion pair using the Fast Gradient Sign Method (FGSM).

[0142]

[0143] Where L is the loss of the downstream model, y is the correct label, and the perturbation magnitude ω is a small constant whose specific value depends on the pixel normalization range of the input image. In this embodiment, the normalization range is [0,1], so ε=0.03 is chosen. The purpose is to add a tiny perturbation that is imperceptible to the naked eye but sufficient to change the model's prediction. The sample pairs formed with the mislabeled sample y' constitute an "adversarial sample pair" and are added to the training data of the S2 discriminator. This forces the discriminator to learn to identify subtle and targeted perturbations, thereby indirectly triggering the generator to produce more discriminative features and improving the robustness of the entire system to perturbations.

[0144] S45. Strategy 3: Adjust the filtering threshold of S3. As described in section S3, S4 will send feedback information to S3 based on the global training state to dynamically adjust the threshold. Thresholds are used to macroscopically control the balance between the quality and diversity of the injected data.

[0145] Through the above multi-strategy linkage, the S5 module ensures that the data augmentation process is always aligned with the real-time needs of the downstream model, forming a self-evolving and continuously optimized intelligent system closed loop. This is the core innovation of this patented method that surpasses static augmentation schemes.

[0146] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.

Claims

1. A generative data augmentation method with dynamic quality feedback, characterized in that, Includes the following steps: S1: Receive a data augmentation request from an external training system, the data augmentation request including a target semantic condition vector and a quantity requirement; determine the category sampling weight distribution based on the target semantic condition vector, sample semantic labels from a predefined category set, encode them, and use them as condition vectors; S2: The random noise vector and the conditional vector are used as common inputs to the generator to generate samples; the discriminator compares the real samples and generated samples in the original training dataset to be enhanced; the generator and discriminator are alternately optimized through the loss function until the optimization condition is met; S3: Perform quality assessment on the samples generated by the generator after S2 optimization to obtain high-quality augmented data; S4: Apply high-quality augmented data to the downstream model, monitor the performance of the downstream model on the validation set, and if the performance is insufficient, determine the adjustment strategy based on the performance, and then return to S1 to generate samples again.

2. The generative data augmentation method with dynamic quality feedback according to claim 1, characterized in that, S1 includes: Construct the probability distribution for the initial class sampling; The category is selected based on the probability distribution; The sampled category labels are encoded into condition vectors; The conditional vector is input into the generator to control the semantic category of the generated samples.

3. The generative data augmentation method with dynamic quality feedback according to claim 1, characterized in that, S3 includes: The authenticity score is obtained by evaluating whether the generated sample approximates the real data in terms of feature distribution; the novelty score is obtained by evaluating whether the generated sample maintains a reasonable difference from the existing training data; the authenticity score and novelty score are compared with the corresponding thresholds, and if both are greater than the corresponding thresholds, they are judged as high-quality augmented data.

4. The generative data augmentation method with dynamic quality feedback according to claim 3, characterized in that, The evaluation of whether the generated samples approximate the real data in terms of feature distribution includes: Calculate category feature centers: For each semantic category k, randomly select from its real training samples. For each sample, features are extracted and the mean is calculated to obtain the feature center of that category. ; Calculate distance: For generated samples with condition label g Extract its features Calculate its distance to the corresponding category center Euclidean distance: Convert to a fraction: Map the Euclidean distance to a fidelity score in the interval [0,1] using a negative exponential function.

5. The generative data augmentation method with dynamic quality feedback according to claim 3, characterized in that, The evaluation of whether the generated samples maintain a reasonable difference from the existing training data, as described in S3, includes: Building a dynamic feature library: Building and maintaining a feature library The feature library Feature vectors containing all real samples and historically qualified augmented samples; Find the nearest neighbor distance: Calculate the features of the generated samples To the feature library The minimum Euclidean distance among all eigenvectors in the set is denoted as . ; Convert to a score: Map the distance to a novelty score in the interval [0,1] using the Sigmoid function.

6. The generative data augmentation method with dynamic quality feedback according to claim 3, characterized in that, The corresponding threshold is obtained by statistical analysis of samples in the original training dataset.

7. The generative data augmentation method with dynamic quality feedback according to claim 1, characterized in that, The analytical performance includes: Analyzing the change in model accuracy before and after data augmentation as an indicator of overall performance change ; The significance threshold is dynamically determined through statistical tests based on the validation set size. ; Calculate the recall change for each category k It identifies weak and declining categories and performs inter-category confusion analysis; the weak category refers to... and <The category of average recall * x; where x = average recall - standard deviation of recall; the declining category refers to... The categories; the inter-class confusion analysis, comparing the confusion matrices before and after enhancement, finds the category pair (i,j) with the largest increase in the absolute difference of the false positive rate, and denoted as the confused category pair.

8. The generative data augmentation method with dynamic quality feedback according to claim 1, characterized in that, The method of determining the adjustment strategy based on performance includes: If overall performance declines The strategy of zero is adopted to achieve dynamic regression control of the training data distribution by jointly adjusting the strength of the generated distribution contraction, the sample selection threshold, the latent space variance, and the proportion of generated data, so that the augmented data distribution can approximate the real data distribution again, thereby restoring the model performance. If the overall performance does not degrade but there are weak / degraded / confused class pairs, then strategy one, strategy two, and strategy three are adopted according to the existence of each class. Strategy one adjusts the conditional input of S1, strategy two adjusts the content of the original training dataset and the generation target of S2, and strategy three adjusts the threshold in the quality evaluation process of S3.

9. The generative data augmentation method with dynamic quality feedback according to claim 8, characterized in that, The strategy zero includes: By adjusting the fidelity constraint parameters of the generated samples This makes the generated samples closer to the real data distribution; Constrain the generator input noise distribution; Control the proportion of augmented data samples to the total number of samples in each training batch of the downstream model.

10. The generative data augmentation method with dynamic quality feedback according to claim 8, characterized in that, Strategy 1 includes: for weak / declining categories, increasing their generation priority; for confused category pairs, generating boundary samples to refine the decision boundary; Strategy 2 includes: for weak categories, extracting features from samples that are misjudged by high loss in the weak category; generating feature-constrained loss based on these features and adding them to the loss function optimized by the regenerator; for confused category pairs, generating adversarial samples from the real samples in them.

Images