A robust target recognition method based on noise analysis, device and storage medium thereof

Abstract

The application discloses a kind of robust target identification method, equipment and its storage medium based on noise analysis. The original data set is established by obtaining original noisy data, the original data set is screened based on K-nearest neighbor method with Jensen-Shannon divergence, and the net data subset and noise data subset are respectively built;The corresponding weight parameter and gradient parameter are obtained in the training process using the network, the response of net data subset and noise data subset is calculated, and the corresponding sensitivity and contribution degree are obtained;Sensitivity and contribution degree are fused, and the overall importance score of parameter is obtained comprehensively;The loss function in the pruning criterion of Taylor expansion is improved by overall importance score, and a pruning model is built.The application fuses the information in feature space and prediction space, accurately divides the clean samples and noise samples in the original data set, and improves the reliability of weight parameter evaluation sample in the pruning process of pruning model.

Description

Technical Field

[0001] Specifically, this invention relates to a robust target recognition method, device, and storage medium based on noise analysis. Background Technology

[0002] In the training and deployment of deep learning models, the presence of data noise often leads to a decline in model performance, especially in large-scale datasets or real-world application scenarios where noisy labels and interfering samples are difficult to avoid. Existing models typically require complex noise modeling or data cleaning mechanisms to handle noisy data. However, these methods often increase memory usage by more than 50% to support complex noise-robust architectures, resulting in reduced batch size and compromised gradient update stability, failing to meet the demands for lightweight and efficient inference. Furthermore, as model size continues to increase, the number of parameters and computational loads increase significantly. Ensuring both noise resistance and computational efficiency when deploying on edge devices or in resource-constrained environments becomes a pressing technical challenge.

[0003] Currently, commonly used model compression and acceleration methods include weight pruning, structured pruning, and quantization. However, most existing pruning methods focus on optimizing parameter redundancy and inference speed, failing to fully consider the impact of data noise on model training and inference, which can easily lead to performance degradation in noisy data scenarios. Therefore, how to enhance noise robustness while maintaining model lightweightness has become an important research direction and technical challenge in this field. Summary of the Invention

[0004] Purpose of the invention: To provide a robust target recognition method, device and storage medium based on noise analysis, which solves the above-mentioned problems existing in the prior art.

[0005] Technical solution: A robust target recognition method based on noise analysis, comprising the following steps:

[0006] The original noisy data was obtained to create an original dataset, and the data in the original dataset were numbered to obtain data labels. The original dataset was then sieved using the K-Nearest Neighbors method combined with Jensen-Shannon divergence. The Jensen-Shannon divergence sieving results were evaluated based on KNN evaluation to separate clean samples and noisy samples, and clean data subsets were constructed for each. and noisy data subset A training model is built based on a subset of noisy data;

[0007] During the training process of the network model, the corresponding weight parameters and gradient parameters are obtained. The net data subset is then calculated using the weight parameters and gradient parameters. and noisy data subset The response yields the corresponding sensitivity and contribution.

[0008] By integrating sensitivity and contribution, and using the pruning criterion based on Taylor expansion, the overall importance score of the parameter is obtained.

[0009] By improving the loss function in the Taylor expansion pruning criterion using the overall importance score, a pruning model is built, utilizing a subset of net data. and noisy data subset The performance of the pruning model is evaluated. If the performance of the pruning model deviates, the network training step is returned for adjustment; otherwise, the current pruning model is used for data identification.

[0010] Preferably, the process of classifying clean and noisy samples in the original dataset using the K-nearest neighbor method is as follows:

[0011] In the feature space, feature vectors are extracted from the trained model. Subset of noisy data The noise data is numbered to obtain the noise label. Obtain feature vectors With noise label Using M in the KNN model knn The algorithm performs classification to obtain the predicted category. The data in the predicted categories are numbered to obtain the predicted labels; the predicted categories are... With noise label Comparison, when predicting category With noise label If consistent, then predict the category. If a sample is identified as clean, then the category is predicted; otherwise, it is considered clean. The sample was identified as noise.

[0012] Preferably, the process of classifying clean and noisy samples in the original dataset using Jensen-Shannon divergence is as follows:

[0013] Within the prediction space, the Jensen-Shannon divergence ranges from [0, 1]. The probability that the data in the original dataset is a clean sample is calculated using the following formula:

[0014] ;

[0015] In the formula: express Divergence calculation function; Indicates the distribution of data labels; Indicates the predicted label distribution; express Divergence; Represents the distribution of data labels and predicted label distribution The average distribution of, where , express The divergence function calculates the probability of clean samples in the original dataset and uses the KNN evaluation results to divide the original dataset into sequences of clean samples. and noisy sample sequences Based on the calculated probability of clean samples For clean sample sequences and noisy sample sequences Arrange them, where clean sequences are ordered by Arranged in descending order, noise sequences are ordered by Sort in ascending order from clean sample sequences Noise sample sequence Select respectively Samples, construct clean subsets and noise subset .

[0016] Preferably, the specific steps for obtaining the corresponding weight parameters and gradient parameters during network training are as follows:

[0017] Obtain the original dataset and divide it into training and test sets. Initialize the training model and perform iterative training using the training set. In each round, perform forward propagation to obtain prediction results, calculate the loss, perform backpropagation, collect gradients, and then update the parameters using an optimization algorithm. After training, the weight parameters are distributed in each connection of the network, and the gradient parameters are stored in the gradient attribute of the weight parameters. The gradient parameters reflect the response of the parameter to the forward propagation dataset, based on the response evaluation function. It is found that the larger the values ​​of the weight parameters and the gradient parameters, the greater the response of the parameters to the corresponding dataset.

[0018] Preferably, the net data subset is used to calculate the weight parameters and gradient parameters. and noisy data subset The response is calculated using the same method, net data subset and noisy data subset The response data obtained is statistically analyzed into a dataset. .

[0019] Preferably, the specific steps for calculating the overall importance score of the parameter are as follows:

[0020] Let the first The first in the layer The parameters are Its impact on datasets The response is denoted as response That is, calculate the first The first in the layer The sensitivity of each parameter to the dataset is further determined by... Reaction number The first in the layer The net data response is calculated by calculating the contribution of each parameter to the entire network model, removing parameters with low contributions and retaining those with high contributions. and noise data response The parameters are calculated using the following formula. The importance score is calculated using the following formula:

[0021] ;

[0022] Wherein: This represents the response calculation function; This represents a balance coefficient used to adjust the relative weights of clean sample responses and noisy sample responses.

[0023] Preferably, for the dataset The response is denoted as The calculation steps are as follows:

[0024] Based on parameters The importance score is approximated by Taylor expansion:

[0025] ;

[0026] In the formula: Indicates parameters Importance score based on Taylor expansion; Represents the loss function; Represents the gradient; Represents a higher-order infinitesimal;

[0027] Based on the Taylor expansion approximation, the expansion is truncated to the first derivative, and higher-order terms are ignored. ,get:

[0028] ;

[0029] Therefore, in this criterion, the gradient information of the loss function directly reflects the parameters. For dataset The degree of response, therefore used as the net data response. and noise data response Approximate value.

[0030] Preferably, the loss function The specific steps for improvement are as follows:

[0031] For the clean sample subset The corresponding clean probability As a weighting factor; for a subset of noisy samples The corresponding noise probability As a weighting factor, the formula for calculating the weighted loss is as follows:

[0032] ;

[0033] The parameters are obtained through the formula for calculating the weighted loss. The final expression for the importance score is:

[0034] ;

[0035] In the formula: Represents the loss function For model weights The gradient; Indicates model weights; This indicates the input sample data.

[0036] Preferred, the optimization objective process of the pruning model is as follows:

[0037] The loss function is calculated using the weight parameters according to the following formula. The changes in the weight parameters are measured by the difference in loss after pruning, i.e., the parameter importance score. The higher the score, the more critical the corresponding weight parameter is to the pruning model; conversely, the lower the score, the more critical the weight parameter is to the pruning model. The calculation formula is as follows:

[0038] ;

[0039] In the formula, Represents the weight parameter vector Norm, i.e., non-zero weight parameter Quantity; This represents the weight parameters after pruning; Indicates the desired sparsity of the pruning model; Indicates being bound by; This represents the difference in loss.

[0040] A computer device for robust target recognition based on noise analysis includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the robust target recognition method based on noise analysis described above.

[0041] A computer-readable storage medium for robust target recognition based on noise analysis, wherein the computer program / instructions, when executed by a processor, implement the steps of the robust target recognition method based on noise analysis described above.

[0042] Beneficial effects: This invention relates to a robust target recognition method, device and storage medium based on noise analysis. By fusing information in the feature space and prediction space, it accurately distinguishes clean samples and noisy samples in the original dataset, improves the reliability of the weight parameter evaluation samples in the pruning process of the pruning model, and reduces the interference of noisy samples on the training and pruning decision of the pruning model.

[0043] Based on the Taylor expansion pruning criterion, the response of each weight parameter and gradient parameter in clean and noisy samples is quantified to improve the recognition of redundant parameters, thereby achieving targeted pruning and retaining the effective knowledge corresponding to clean samples.

[0044] In the parameter importance assessment, dynamic weight allocation is introduced to enhance the influence of clean samples on pruning decisions, thereby improving the effectiveness and robustness of pruning. Attached Figure Description

[0045] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0046] like Figure 1 As shown, the present invention provides a technical solution: a robust target recognition method, device, and storage medium based on noise analysis, comprising the following steps:

[0047] The original noisy data is obtained to construct the original dataset, and the data in the original dataset are numbered to obtain data labels. In the feature space, the K-Nearest Neighbors method is used to classify clean and noisy samples in the original dataset. Jensen-Shannon divergence is then used in the prediction space to classify clean and noisy samples in the original dataset. Finally, the Jensen-Shannon divergence screening results are evaluated based on KNN evaluation, separating clean and noisy samples, and constructing clean data subsets for each. and noisy data subset Based on a subset of noisy data The process of building and training a model, including classifying clean and noisy samples in the original dataset using the K-nearest neighbor method, is as follows:

[0048] In the feature space, feature vectors are extracted from the trained model. Subset of noisy data The noise data is numbered to obtain the noise label. Obtain feature vectors With noise label Using M in the KNN model knn The algorithm performs classification to obtain the predicted category. The data in the predicted categories are numbered to obtain the predicted labels; the predicted categories are... With noise label Comparison, when predicting category With noise label If consistent, then predict the category. If a sample is identified as clean, then the category is predicted; otherwise, it is considered clean. The process of classifying clean and noisy samples in the original dataset using Jensen-Shannon divergence, after identifying them as noise samples, is as follows:

[0049] Within the prediction space, the Jensen-Shannon divergence ranges from [0, 1]. The probability that the data in the original dataset is a clean sample is calculated using the following formula:

[0050] ;

[0051] In the formula: express Divergence calculation function; Indicates the distribution of data labels; D represents the predicted label distribution; KL Indicates KL divergence; Represents the distribution of data labels and predicted label distribution The average distribution of, where , express The divergence function calculates the probability of clean samples in the original dataset and uses the KNN evaluation results to divide the original dataset into sequences of clean samples. and noisy sample sequences Based on the calculated probability of clean samples For clean sample sequences and noisy sample sequences Arrange them, where clean sequences are ordered by Arranged in descending order, noise sequences are ordered by Sort in ascending order from clean sample sequences Noise sample sequence Select respectively Samples, construct clean subsets and noise subset By fusing information from the feature space and the prediction space, clean and noisy samples in the original dataset can be accurately separated.

[0052] During the training process of the network model, the corresponding weight parameters and gradient parameters are obtained. The net data subset is then calculated using the weight parameters and gradient parameters. and noisy data subset The responses are analyzed to obtain corresponding sensitivity and contribution values. By fusing these sensitivity and contribution values ​​and applying a Taylor expansion-based pruning criterion, an overall importance score for the parameters is obtained. Based on this overall importance score, the loss function in the Taylor expansion-based pruning criterion is improved, and a pruning model is constructed using a subset of net data. and noisy data subset The performance of the pruning model is evaluated. If the performance of the pruning model deviates, the network training step is repeated for adjustment. Otherwise, the current pruning model is used for data identification. First, the clean and noisy samples in the original dataset are accurately divided. Based on the Taylor expansion pruning criterion, the response of each weight parameter and gradient parameter to clean and noisy samples is quantified to improve the identification of redundant parameters, thereby achieving targeted pruning while retaining the effective knowledge corresponding to clean samples. In the parameter importance evaluation, dynamic weight allocation is introduced to enhance the influence of clean samples on pruning decisions, thereby improving the effectiveness and robustness of pruning.

[0053] In a further embodiment, the specific steps for obtaining the corresponding weight parameters and gradient parameters using the network during the model training process are as follows:

[0054] Obtain the original dataset and divide it into training and test sets. Initialize the training model and perform iterative training using the training set. In each round, perform forward propagation to obtain prediction results, calculate the loss, perform backpropagation, collect gradients, and then update the parameters using an optimization algorithm. After training, the weight parameters are distributed in each connection of the network, and the gradient parameters are stored in the gradient attribute of the weight parameters. The gradient parameters reflect the response of the parameter to the forward propagation dataset, based on the response evaluation function. It is found that the larger the values ​​of the weight parameters and gradient parameters, the greater the response of the parameters to the corresponding dataset. Specifically, the weight parameters and gradient parameters are used to calculate the net data subset. and noisy data subset The response is calculated using the same method, net data subset and noisy data subset The response data obtained is statistically analyzed into a dataset. For the dataset The response is denoted as The calculation steps are as follows:

[0055] Based on parameters The importance score is approximated by Taylor expansion:

[0056] ;

[0057] In the formula: Indicates parameters Importance score based on Taylor expansion; Represents the loss function; Represents the gradient; Represents a higher-order infinitesimal, where the loss function The specific steps for improvement are as follows:

[0058] For the clean sample subset The corresponding clean probability As a weighting factor; for a subset of noisy samples The corresponding noise probability As a weighting factor, the formula for calculating the weighted loss is as follows:

[0059] ;

[0060] The parameters are obtained through the formula for calculating the weighted loss. The final expression for the importance score is:

[0061] ;

[0062] In the formula: Represents the loss function right The gradient; Indicates model weights; This indicates the input sample data.

[0063] Based on the Taylor expansion approximation, the expansion is truncated to the first derivative, and higher-order terms are ignored. ,get:

[0064] ;

[0065] Therefore, in this criterion, the gradient information of the loss function directly reflects the parameters. For dataset The degree of response, therefore used as the net data response. and noise data response Approximate value.

[0066] In a further embodiment, the specific steps for calculating the overall importance score of the parameter are as follows:

[0067] Let the first The first in the layer The parameters are Its impact on datasets The response is denoted as response That is, calculate the first The first in the layer The sensitivity of each parameter to the dataset is further determined by... Reaction number The first in the layer The net data response is calculated by calculating the contribution of each parameter to the entire network model, removing parameters with low contributions and retaining those with high contributions. and noise data response The parameters are calculated using the following formula. The importance score is calculated using the following formula:

[0068] ;

[0069] Wherein: This represents the response calculation function; This represents a balance coefficient used to adjust the relative weights of clean sample responses and noisy sample responses.

[0070] In a further embodiment, the optimization objective process of the pruning model is as follows:

[0071] The loss function is calculated using the weight parameters according to the following formula. The changes in the weight parameters are measured by the difference in loss after pruning, i.e., the parameter importance score. The higher the score, the more critical the corresponding weight parameter is to the pruning model; conversely, the lower the score, the more critical the weight parameter is to the pruning model. The calculation formula is as follows:

[0072] ;

[0073] In the formula, Represents the weight parameter vector Norm, i.e., non-zero weight parameter Quantity; This represents the weight parameters after pruning; Indicates the desired sparsity of the pruning model; Indicates being bound by; This represents the difference in loss.

[0074] A computer device for robust target recognition based on noise analysis includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the robust target recognition method based on noise analysis described above.

[0075] A computer-readable storage medium for robust target recognition based on noise analysis, wherein the computer program / instructions, when executed by a processor, implement the steps of the robust target recognition method based on noise analysis described above.

[0076] Based on the above solution, the following implementation method is given:

[0077] In this embodiment, the synthetic datasets Cifar100N and Cifar80N are used, and the verification experiments are conducted using the real noise datasets Web-Aircraft, Web-Bird, and Web-Car.

[0078] We synthesized a damaged dataset using CIFAR100N and CIFAR80N, and a realistic noise dataset using Web-Aircraft, Web-Bird, and Web-Car. CIFAR100N, derived from CIFAR100, was used to construct a closed-set noise dataset. Symmetric and asymmetric noise were then introduced and studied within this closed-set noise dataset. Similarly, the open-set noise dataset CIFAR80N was constructed, where the last 20 classes of CIFAR100 were considered out-of-distribution categories. Furthermore, evaluation experiments were conducted on real-world datasets, including Web-Aircraft, Web-Bird, and Web-Car. Training samples were collected using a web image search engine, and real-world noise labels were incorporated to construct these datasets.

[0079] To evaluate the performance of this method, test accuracy was used as the evaluation metric. To reduce the impact of randomness and ensure fair comparison, the average result of three repeated runs under the same settings was recorded.

[0080] For the constructed closed-set noisy dataset, pruning and fine-tuning were performed on ResNet-18. Pruning was performed after training on the specified network to a moderate overfit using the original dataset for a specified number of epochs, with pruning rates of 0.2 and 0.25, respectively. After pruning, the model was fine-tuned to maintain performance. The pruned model was fine-tuned for 15 epochs using a learning rate of 0.01, a batch size of 128, and the SGD optimizer. Other hyperparameters were set as follows: , , This is the default value. For the real dataset, a ResNet-50 model trained using NPN was pruned. The pruning ratio was set to 0.1. After pruning, fine-tuning was performed according to the robust training process of NPN. Fine-tuning was performed in two phases: first, a warm-up training of 10 phases with a fixed learning rate of 0.005 was performed, followed by robust training of 120 phases using a cosine annealing learning rate plan starting at 0.008. , , Set as the default value.

[0081] To evaluate the effectiveness of the inventive method, it was compared with the following state-of-the-art (SOTA) methods: L2, Taylor, FPGM, LAMP, Group L2, Group Taylor, and Class aware. To ensure a fair comparison, all the comparison methods were reimplemented based on Torch-Pruning.

[0082] Extensive experiments were conducted on closed-set noisy datasets to validate the effectiveness of the proposed method. In these experiments, the baseline ResNet-18 model used for pruning had already experienced performance degradation due to overfitting to noisy labels. All pruning methods pruned the model with the same pruning rate to remove redundant parameters. After fine-tuning, the model performance partially recovered. Thanks to the noise-aware pruning criterion designed in this invention, our method achieved the highest performance recovery among all methods. Compared to the baseline ResNet-18, the post-fine-tuned model exhibited superior performance, demonstrating that our method effectively removed parameters that were overfitted to noisy labels.

[0083] Table 1 Comparative experiment on Cifar 100N using a pruning rate of 0.2.

[0084]

[0085] Table 2 Comparative experiments on Cifar 100N using a pruning rate of 0.25.

[0086]

[0087] Tables 1 and 2 present experimental comparisons on CIFAR100N, where the baseline ResNet-18 model was pruned at rates of 0.2 and 0.25, respectively. At a pruning rate of 0.2, the number of parameters was reduced by 36.2% and the computational cost by 35.3%. At a pruning rate of 0.25, the number of parameters was reduced by 43.6% and the computational cost by 42.4%. These comparisons demonstrate that our method consistently outperforms other methods on various dataset settings with different noise types and rates.

[0088] Table 3 Comparative experiments on Cifar 80N using a pruning rate of 0.2.

[0089]

[0090] Table 4 Comparative experiments on Cifar 80N using a pruning rate of 0.25.

[0091]

[0092] Tables 3 and 4 extend the comparison of CIFAR80N under the same pruning rates of 0.2 and 0.25, respectively. Compared with other methods, our method again achieves superior or competitive results. This further confirms the effectiveness and generalization ability of the method. Notably, the performance improvement of our method becomes more pronounced with increasing noise rate. In the most challenging case, when the noise rate reaches 80%, our method demonstrates significant superiority.

[0093] Table 5 Comparative experiments on the Web-Aircraft, Web-Bird, and Web-Car datasets.

[0094]

[0095] Table 5 evaluates our method on three real-world web datasets using a state-of-the-art (SOTA) noise-robust NPN model. With a pruning rate of 0.1, the number of parameters is reduced by 16.2%, and the computational cost by 19.1%. The model used for pruning is a state-of-the-art robust model specifically optimized for noisy data, making it difficult to improve accuracy through pruning. Thanks to the noise-aware pruning design of this invention, our method still outperforms all other pruning strategies. Notably, our method even slightly improves the accuracy of the baseline NPN on the Web-Aircraft and Web-Car datasets, further demonstrating its superiority.

[0096] In summary, the pruning method proposed in this invention is an effective method. This method ensures that parameters overfitting to noisy data are removed during the pruning process, while retaining key parameters that capture effective features from clean data. Therefore, it is superior to other pruning methods in maintaining model performance stability.

[0097] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A robust target recognition method based on noise analysis, characterized in that, The method comprises the following steps: The original noisy data is acquired to establish an original data set, and the data in the original data set is numbered to obtain a data label, the original data set is screened based on a K-neighbor method and a Jensen-Shannon divergence, and the screening result of the Jensen-Shannon divergence is evaluated based on KNN evaluation, to divide clean samples and noise samples, and build a clean data subset and a noise data subset respectively , and a training model is built based on the noise data subset; The corresponding weight parameters and gradient parameters are obtained in the training process of the training model by using the network, and the responses of the net data subset and the noise data subset are calculated by using the weight parameters and the gradient parameters The corresponding sensitivity and contribution degree are obtained The calculation of the responses of the net data subset and the noise data subset by using the weight parameters and the gradient parameters adopts the same calculation method The responses of the net data subset and the noise data subset The response data statistics of the responses of the net data subset and the noise data subset The response data statistics of the responses of the net data subset and the noise data subset are the data set The specific calculation steps of the overall importance score of the parameters are as follows: Let the th parameter in the th layer be , the response of the data set is denoted as response , that is, the sensitivity of the th parameter in the th layer to the data set is calculated, further by reaction the th parameter in the th layer contributes to the entire network model, the parameters with low contribution are pruned while the parameters with high contribution are retained, and the net data response and the noise data response are calculated, the importance score of the parameter is calculated by the following formula, the formula is as follows: ； wherein, in the formulae: represents a response calculation function; represents a balancing coefficient for adjusting the relative weight of the clean sample response and the noise sample response; where the response to the data set is denoted by The calculation steps are as follows: Based on the parameter The importance score is approximated by a Taylor expansion: ； wherein: denotes a parameter importance score based on Taylor expansion; denotes a loss function denotes a gradient denotes a higher order infinitesimal Based on the Taylor expansion approximation, the expansion is truncated to the first derivative and the higher order terms are ignored , which gives ； Therefore, in the pruning criterion of the Taylor expansion, the gradient information of the loss function directly reflects the parameter response to the data set , so it is used as an approximation of the net data response and the noise data response ; The sensitivity and the contribution degree are fused, and the overall importance score of the parameters is obtained based on a Taylor expansion pruning criterion; By improving the loss function in the pruning criterion of Taylor expansion, a pruning model is built, and the net data subset and the noise data subset The performance of the pruning model is evaluated, and when the performance of the pruning model deviates, the network training step is returned for adjustment, otherwise, the current pruning model is used for data recognition.

2. The robust target recognition method based on noise analysis according to claim 1, characterized in that, The process of classifying the original data set into clean samples and noise samples by using the K-neighbor method is as follows: extracting feature vectors from the training model in the feature space , obtaining noise labels of the noise data in the noise data subset , obtaining noise labels of the noise data in the noise data subset , obtaining feature vectors , obtaining feature vectors , using M knn algorithm in the KNN model to classify, obtaining predicted classes , obtaining predicted labels by numbering the data in the predicted classes; comparing the predicted classes with the noise labels , when the predicted classes are consistent with the noise labels , the predicted classes are identified as clean samples, otherwise, the predicted classes are identified as noise samples.

3. The robust target recognition method based on noise analysis according to claim 1, characterized in that, The process of classifying the original data set into clean samples and noise samples by using the Jensen-Shannon divergence is as follows: In the prediction space, the value range of the Jensen-Shannon divergence is [0, 1], and the probability that the data in the original data set is a clean sample is calculated by the following calculation formula: ； In the formula: express Divergence calculation function; Indicates the distribution of data labels; Indicates the predicted label distribution; express Divergence; Represents the distribution of data labels and predicted label distribution The average distribution of, where , express The divergence function calculates the probability of clean samples in the original dataset and uses the KNN evaluation results to divide the original dataset into sequences of clean samples. and noisy sample sequences Based on the calculated probability of clean samples For clean sample sequences and noisy sample sequences Arrange them, where clean sequences are ordered by Arranged in descending order, noise sequences are ordered by Sort in ascending order from clean sample sequences Noise sample sequence Select respectively Samples, construct clean subsets and noise subset .

4. The robust target recognition method based on noise analysis according to claim 1, characterized in that, The specific steps of obtaining the corresponding weight parameters and gradient parameters of the network in the training process are as follows: The original data set is obtained, the original data set is divided into a training set and a test set, a training model is initialized, and the training set is used for cyclic training, a prediction result is obtained by forward propagation before each round, a loss is calculated and back propagation is performed, gradients are collected, and parameters are updated using an optimization algorithm, and a result model is obtained after training, wherein the weight parameters are distributed in each connection of the network, and the gradient parameters are saved in the gradient attribute of the weight parameters, wherein the gradient parameters reflect the response of the parameters to the forward propagation of the data set, and a response evaluation function is obtained according to the response The greater the weight parameter value and the greater the gradient parameter value, the greater the response of the parameters to the corresponding data set.

5. The robust target recognition method based on noise analysis according to claim 1, characterized in that, Loss function The improved specific steps are as follows: For the clean sample subset its corresponding clean probability as a weighting factor; for the noisy sample subset its corresponding noisy probability The calculation formula of the weighted loss is as follows: ； The parameters are obtained by the calculation formula of the weighted loss The final expression of the importance score is: ； where: represents a loss function the gradient of the model weights ; represents the model weights; represents the input sample data.

6. The robust target recognition method based on noise analysis according to claim 1, characterized in that, The optimization target process of the pruning model is as follows: The weight parameters are used to calculate the loss function by the following formula The importance of the weight parameters is measured by the loss difference after pruning, i.e., the parameter importance score The higher the score is, the more critical the weight parameter is to the pruned model, and vice versa, the weight parameter is pruned, wherein the calculation formula is as follows: ； wherein denotes a weight parameter vector denotes a norm, i.e. a number of non-zero weight parameters denotes a pruned weight parameter denotes a pruned weight parameter denotes a desired sparsity of the pruned model denotes a constraint denotes a loss difference 7. A computer device for robust target identification of noise analysis, characterized in that, The memory and the processor, the memory stores the computer program, the computer program is executed by the processor, so that the processor executes the steps of the robust target identification method of noise analysis in any one of claims 1-6.

8. A computer-readable storage medium for robust target identification of noise analysis, characterized in that, Executable instructions are stored thereon, which are executed by a processor to implement the steps of the robust target identification method of noise analysis in any one of claims 1-6.

Images