Robust noise tag domain adaptive method based on confidence coefficient and information entropy

Abstract

The invention discloses a robust noise tag domain adaptive method based on confidence and information entropy. The method specifically comprises the following steps: preparing data in a double-domain data set of a source domain and a target domain; in the preheating stage, the model is trained by adopting source domain labeled data; confidence modulation adaptive learning is carried out; a confusion alignment mechanism based on entropy guidance; establishing a joint optimization loss function; and outputting an adaptive result. By introducing a joint optimization strategy of confidence adjustment learning and entropy guide confusion alignment, the learning intensity is dynamically adjusted, and the inter-class confusion is reduced, so that the problems caused by noise labels and cross-domain distribution differences are effectively solved, and the robustness and stability of field self-adaption are improved; under the condition that there is no label in the target domain, optimization is carried out by means of prediction confidence and entropy information, and the target domain classification accuracy under the condition of high noise label proportion and unbalanced category distribution is improved.

Description

Technical Field

[0001] The present invention belongs to the field of noise label processing, and in particular relates to a robust noise label domain adaptation method based on confidence and information entropy. Background Art

[0002] Deep learning models face performance degradation when processing cross-domain data. Even if a model performs well in the training environment, its performance often degrades significantly in the deployed environment due to the shift in distribution. Unsupervised domain adaptation techniques have emerged and made significant progress, primarily addressing the distribution discrepancy between the source and target domains through methods such as feature alignment, adversarial training, and pseudo-label generation. However, most of these methods rely on fixed noise processing strategies, lack dynamic adjustment capabilities, and have difficulty adapting to different noise distributions. Most existing unsupervised domain adaptation methods assume that the source domain labels are clean, an assumption that is difficult to hold in real-world scenarios. Furthermore, these methods focus more on sample-level noise processing and lack the ability to perform category-level alignment and boundary optimization, which can easily lead to confusion between categories in the target domain and affect the accuracy of classification tasks. More importantly, noisy labels can be amplified by the model during training, leading to error accumulation and significantly reducing the adaptation performance of the target domain under high-noise conditions. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a robust noisy label domain adaptation method based on confidence and information entropy. This method effectively prevents the model from overfitting on noisy labels by dynamically adjusting the sample-level learning intensity. At the same time, it guides alignment through category-level entropy, optimizes the relationship between categories, maintains clear decision boundaries, and improves the overall robustness and accuracy of the adaptation process.

[0004] In order to solve the above technical problems, the present invention is implemented in the following ways:

[0005] A robust noise label domain adaptation method based on confidence and information entropy specifically includes the following steps:

[0006] S1. Prepare data in the source domain and target domain dual-domain datasets;

[0007] S2, the warm-up phase, where the model is trained using labeled data from the source domain;

[0008] S3, confidence modulation adaptive learning;

[0009] S4, entropy-guided obfuscation alignment mechanism;

[0010] S5. Establish a joint optimization loss function;

[0011] S6. Output the adaptive result.

[0012] Furthermore, the specific method of step S1 is as follows:

[0013] The dual-domain dataset D={I s , I t}Including source domain data I s and target domain data I t , the source domain data contains N s Sample pairs represents the i-th source domain sample, Represents the corresponding category label, the category label value range is {1,2,...,C}, the target domain data contains N t samples

[0014] The original data is input into the deep convolutional neural network model, and the input image data is normalized, its pixel values are scaled to the range of [0, 1], and standardized according to the ImageNet pixel value mean μ = [0.485, 0.456, 0.406] and pixel value standard deviation σ = [0.229, 0.224, 0.225]. The expression is as follows:

[0015]

[0016] Among them, x norm represents the normalized image data, and x represents the pixel value of the original image data;

[0017] A certain proportion of symmetric noise and asymmetric noise are injected into the source domain data, and ResNet-50 is used as the basic backbone network to test the robustness under different noise intensities.

[0018] Furthermore, the specific method of step S2 is as follows:

[0019] After data preparation is completed, the early learning characteristics of deep neural networks are used to quickly establish feature representation in the initial stage. The model uses source domain data samples to Optimize the loss function, the specific expression is as follows:

[0020]

[0021] in, represents the classification loss of source domain samples (such as cross entropy loss), N s represents the total number of source domain sample pairs;

[0022] The warm-up phase lasts for T warmDuring the training process, the model performance is evaluated on the source domain data subset every several steps. After the warm-up is completed, the model parameters are saved to provide a good initialization for the subsequent training stages.

[0023] Furthermore, the specific method of step S3 is as follows:

[0024] Through the domain shared network f = [f1, ..., f C ] to predict the dual domain data, where f i Represents the output of the i-th category, and the network prediction result is Contains source domain prediction probability and target domain prediction probability The predicted probability is calculated by the softmax function, and the specific expression is as follows:

[0025]

[0026] in, Represents source domain samples The network output score on the jth class, Represents source domain samples The network output score on the kth class, Represents the target domain sample The network output score on the jth class, Represents the target domain sample The network output score on the kth class, τ represents the temperature parameter, which is used to adjust the output smoothness of the Softmax function;

[0027] Using the model parameters and feature representations obtained in the warm-up phase, we introduce a mechanism to dynamically adjust the learning intensity of samples based on prediction confidence and construct an adaptive loss function based on prediction confidence. The expression is as follows:

[0028]

[0029] in, represents the adaptive loss of a single sample, represents the predicted probability of the i-th source domain sample for the j-th class, q represents the adaptive modulation parameter, which belongs to the (0,1) interval;

[0030] For the adaptive loss function, the gradient expression is as follows:

[0031]

[0032] In the case of low prediction confidence, the gradient expression When q-1 is less than 0, the gradient amplitude will be automatically reduced, thereby limiting the negative impact that noise samples may bring. At the same time, The item remains relatively stable, ensuring that the model still has a certain learning ability; when the prediction confidence is high, The term will produce a moderate learning signal, while The term approaches 0, which prevents the model from making overconfident predictions. This natural gradient adjustment mechanism enables the model to fully learn reliable samples while automatically reducing the impact of noise samples.

[0033] The overall loss is calculated as follows:

[0034]

[0035] Here, N represents the batch size.

[0036] Furthermore, the specific method of step S4 is as follows:

[0037] In order to reduce category confusion and maintain a clear decision boundary, an entropy-guided confusion alignment mechanism is proposed. The specific method is as follows:

[0038] For the target domain sample, the cross entropy value is calculated using the target domain prediction probability. The specific expression is as follows:

[0039]

[0040] in, Represents the target domain sample The predicted probability of category j, C represents the total number of categories;

[0041] Based on the entropy value of the target sample, an entropy-guided weight generation function is designed to capture the uncertainty of the sample. The specific expression is as follows:

[0042]

[0043] Among them, α represents the parameter that controls the steepness of the weight distribution, e i Represents the target domain sample The entropy value of

[0044] At the same time, in order to capture the relationship between categories, the predicted probability of the target sample is used to generate the category alignment matrix, which is expressed as follows:

[0045]

[0046] Among them, diag(w) means according to w i The resulting diagonal matrix, A ij represents the element in the i-th row and j-th column of matrix A;

[0047] By performing row normalization on the category alignment matrix A, the normalized category alignment matrix is obtained The expression is as follows:

[0048]

[0049] In order to reduce the uncertainty between categories, a category confusion loss function based on entropy is designed. The specific expression is as follows:

[0050]

[0051] in, Representation matrix The trace represents the contribution of correct classification. By minimizing this loss function, the confusion between categories can be effectively reduced and a clear decision boundary can be maintained.

[0052] Furthermore, the specific method of step S5 is as follows:

[0053] Through the confidence-controlled adaptive learning loss and the entropy-guided confusion alignment loss, robustness optimization in cross-domain adaptation from the source domain to the target domain is achieved, and the expression of the joint optimization loss function is obtained as follows:

[0054] L total =L CMAL +λL EGCA

[0055] Where λ represents the balance parameter, which is used to adjust L CMAL and L EGCA The relative importance in the joint optimization can be adjusted according to the noise ratio and category confusion level of the specific task to achieve the best adaptation effect.

[0056] Furthermore, the specific method of step S6 is as follows:

[0057] The classification accuracy is used to evaluate the overall prediction accuracy of the model on the target domain samples. The classification accuracy is expressed as follows:

[0058]

[0059] Among them, N t represents the total number of target domain samples, y t represents the true label of the target domain sample, represents the indicator function, Represents the predicted label of the target domain sample; if the prediction result Equal to the true label y t , the indicator function takes the value 1, otherwise it is 0.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] The present invention introduces a joint optimization strategy of confidence-modulated learning (CMAL) and entropy-guided confusion alignment (EGCA) to dynamically adjust learning intensity and reduce inter-category confusion, effectively solving the problems caused by noisy labels and cross-domain distribution differences, and improving the robustness and stability of domain adaptation. When the target domain is unlabeled, it relies on prediction confidence and entropy information for optimization, significantly improving the target domain classification accuracy under conditions of a high proportion of noise labels and unbalanced category distribution, ensuring the high applicability of the model. At the same time, it does not rely on the real labels of the target domain or additional labeling process, and belongs to the category of unsupervised domain adaptation, which greatly reduces manpower and time costs, while improving the efficiency and accuracy of cross-domain migration. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0063] The specific implementation of the present invention will be further described in detail below with reference to the accompanying drawings and specific examples.

[0064] like Figure 1 As shown in FIG, a robust noise label domain adaptation method based on confidence and information entropy includes the following steps:

[0065] Experiments were conducted based on the VisDA dataset, a large-scale cross-domain classification dataset consisting of two domains: the Synthetic (source domain), which consists of synthetically rendered images with significant visual differences between categories; and the Real (target domain), which consists of real-world images with a significant distribution difference from the source domain. The VisDA dataset covers 12 categories, including common objects in life such as airplanes, bicycles, and cars. It is a classic platform for validating the adaptability of cross-domain classification methods.

[0066] S1. Prepare the data in the source and target domain datasets. The specific method is as follows:

[0067] The dual-domain dataset D={I s , I t}Including source domain data I s and target domain data I t , the source domain data contains N s Sample pairs represents the i-th source domain sample, Indicates the corresponding category label, the category label value range is {1, 2, .., C}, the target domain data contains N t samples

[0068] The original data is input into the deep convolutional neural network model, and the input image data is normalized, its pixel values are scaled to the range of [0, 1], and standardized according to the ImageNet pixel value mean μ = [0.485, 0.456, 0.406] and pixel value standard deviation σ = [0.229, 0.224, 0.225]. The expression is as follows:

[0069]

[0070] Among them, x norm represents the normalized image data, and x represents the pixel value of the original image data;

[0071] A certain proportion of symmetric noise or asymmetric noise is injected into the source domain data, and the target domain data is completely unlabeled and has significant differences in feature distribution from the source domain data.

[0072] Symmetrical noise: The labels of the source domain data are randomly perturbed according to a uniform distribution. For example, under the condition of 60% symmetric noise, 40% of the sample labels are randomly replaced with other categories, while the remaining 60% of the sample labels remain correct. This noise assumes that the degree of confusion of all categories is the same.

[0073] Asymmetric noise: Asymmetric noise simulates a more realistic label noise pattern, where sample labels are primarily misassigned to semantically similar categories. For example, a sample from the "Car" category might be mislabeled as a "Truck" category. This type of noise assumes a certain degree of semantic correlation between categories, making it more challenging.

[0074] ResNet-50 is used as the basic backbone network. Three symmetric noise ratios of 40%, 60% and 80%, and two asymmetric noise ratios of 20% and 40% are set to test the robustness under different noise intensities.

[0075] In the warm-up phase, the model is trained using labeled data from the source domain. The specific method is as follows:

[0076] After data preparation is completed, the early learning characteristics of deep neural networks are used to quickly establish feature representation in the initial stage. The model uses source domain data samples to Optimize the loss function, the specific expression is as follows:

[0077]

[0078] in, represents the classification loss of source domain samples (such as cross entropy loss), N s represents the total number of source domain sample pairs;

[0079] The warm-up phase lasts for Twarm An epoch indicates that the model completes a complete traversal of the entire training dataset. During the training process, the model performance is evaluated on a subset of the source domain data every several steps. After the warm-up is completed, the model parameters are saved to provide a good initialization for subsequent training stages.

[0080] S3, confidence modulation adaptive learning, the specific method is as follows:

[0081] Through the domain shared network f = [f1, ..., f C}Predict the dual domain data, where f i Represents the output of the i-th category, and the network prediction result is Contains source domain prediction probability and target domain prediction probability The predicted probability is calculated by the softmax function, and the specific expression is as follows:

[0082]

[0083] in, Represents source domain samples The network output score on the jth class, Represents source domain samples The network output score on the kth class, Represents the target domain sample The network output score on the jth class, Represents the target domain sample The network output score on the kth class, τ represents the temperature parameter, which is used to adjust the output smoothness of the Softmax function;

[0084] The model has learned a certain degree of source domain feature distribution, but there may still be interference from noise labels and inconsistency in inter-domain features. To further address these problems, the confidence-modulated adaptive learning phase uses the model parameters and feature representations obtained in the warm-up phase to introduce a mechanism to dynamically adjust the learning intensity of samples based on prediction confidence, thereby improving the model's cross-domain adaptability. This mechanism naturally adjusts the learning intensity according to the prediction confidence by designing a structurally adaptive loss function. For high-confidence predictions, appropriate learning signals are generated while preventing overconfident predictions; for low-confidence predictions (which may come from noise labels), their gradient amplitudes are naturally reduced to limit their negative impact. This design does not require explicit setting of the confidence threshold and can automatically adapt to different noise patterns and domain relationship changes. The present invention constructs an adaptive loss function based on prediction confidence, which is expressed as follows:

[0085]

[0086] in, represents the adaptive loss of a single sample, represents the predicted probability of the i-th source domain sample for the j-th class, q represents the adaptive modulation parameter, which belongs to the (0,1) interval;

[0087] For the adaptive loss function, the gradient expression is as follows:

[0088]

[0089] In the case of low prediction confidence, the gradient expression When q-1 is less than 0, the gradient amplitude will be automatically reduced, thereby limiting the negative impact that noise samples may bring. At the same time, The item remains relatively stable, ensuring that the model still has a certain learning ability; when the prediction confidence is high, The term will produce a moderate learning signal, while The term approaches 0, which prevents the model from making overconfident predictions. This natural gradient adjustment mechanism enables the model to fully learn reliable samples while automatically reducing the impact of noise samples.

[0090] The overall loss is calculated as follows:

[0091]

[0092] Here, N represents the batch size.

[0093] S4. Entropy-guided obfuscation alignment mechanism. The specific method is as follows:

[0094] During domain adaptation, class confusion between samples is a problem that requires special attention. To reduce class confusion and maintain a clear decision boundary, we propose an entropy-guided confusion alignment mechanism that systematically reduces inter-class uncertainty through prediction uncertainty analysis. The specific method is as follows:

[0095] For the target domain sample, the cross entropy value is calculated using the target domain prediction probability. The specific expression is as follows:

[0096]

[0097] in, Represents the target domain sample The predicted probability of category j, C represents the total number of categories; the predicted entropy reflects the model's confidence in the sample classification, and a higher entropy value indicates that the model has greater uncertainty in the sample.

[0098] Based on the entropy value of the target sample, an entropy-guided weight generation function is designed to capture the uncertainty of the sample. The specific expression is as follows:

[0099]

[0100] Among them, α represents the parameter that controls the steepness of the weight distribution, e i Represents the target domain sample The entropy value of ; This function ensures that low entropy samples are given higher weights, thereby prioritizing the optimization of samples with high confidence in the model, while giving lower weights to high entropy samples (which may have category confusion) to reduce their influence.

[0101] In order to capture the relationship between categories, the predicted probability of the target sample is used to generate the category alignment matrix, which is expressed as follows:

[0102]

[0103] Among them, diag(w) means according to w i The resulting diagonal matrix, A ij represents the element in the i-th row and j-th column of matrix A;

[0104] By performing row normalization on the category alignment matrix A, the normalized category alignment matrix is obtained The expression is as follows:

[0105]

[0106] In order to reduce the uncertainty between categories, a category confusion loss function based on entropy is designed. The specific expression is as follows:

[0107]

[0108] in, Representation matrix The trace represents the contribution of correct classification. By minimizing this loss function, the confusion between categories can be effectively reduced and a clear decision boundary can be maintained.

[0109] S5. Establish a joint optimization loss function. The specific method is as follows:

[0110] In step S3, the source domain data is processed and a CMAL (Confidence-Modulated Adaptive Learning) loss is designed for sample-level optimization. In step S4, the EGCA (Entropy-Guided Confusion Alignment) loss is introduced to align the category distributions of the source and target domains by minimizing the uncertainty (entropy value) of the category distributions. To achieve robustness optimization in cross-domain adaptation from the source domain to the target domain, the two losses are combined to obtain a joint optimization loss function. The core idea of this loss function is to balance sample reliability and category decision clarity at the same time, and to jointly optimize the joint loss function using labeled data from the source domain and unlabeled data from the target domain. The expression of the joint optimization loss function is as follows:

[0111] L total =L CMAL +λL EGCA

[0112] Among them, λ represents the balance parameter, which is used to adjust L CMAL and L EGCA The relative importance of L in joint optimization can be adjusted according to the noise ratio and category confusion of the specific task to achieve the best adaptation effect. Each iteration extracts a batch of data from the source domain data and the target domain data, and calculates L CMAL and L EGCA , based on the joint optimization loss function, the model parameters are updated through back propagation; the model performance is evaluated on the validation set of the target domain every several steps to observe the changes in classification accuracy.

[0113] S6. Output the adaptive results. The specific method is as follows:

[0114] Use the joint loss function in step S5 to train the model, perform forward propagation on the target domain test set, and predict the category of each sample. Record the prediction results on the target domain test set and the true label y t , the classification accuracy is calculated based on the prediction results, and the classification accuracy is used for evaluation to measure the overall prediction accuracy of the model on the target domain samples. The expression of the classification accuracy is as follows:

[0115]

[0116] Among them, N t represents the total number of target domain samples, y t Represents the true label of the target domain sample, which is the target domain data with the correct label output by the model. represents the indicator function, Represents the predicted label of the target domain sample; if the prediction result Equal to the true label y t , the indicator function takes the value of 1, otherwise it is 0. This indicator can intuitively reflect the classification effect of the model in the target domain. The higher the accuracy, the better the performance of the model in cross-domain adaptation.

[0117] In this example, the classification accuracy rates were 72.3%, 71.5%, and 67.2% under symmetric noise levels of 40%, 60%, and 80%, respectively; and 70.8% and 67.4% under asymmetric noise levels of 20% and 40%, respectively. Comparisons show that the classification accuracy of this method significantly outperforms existing methods under various noise conditions, demonstrating its effectiveness in handling noisy labels and cross-domain adaptation.

[0118] The above description is merely an embodiment of the present invention. It is stated again that, for a person skilled in the art, several improvements can be made to the present invention without departing from the principles of the present invention, and these improvements are also included in the scope of protection of the claims of the present invention.

Claims

1. A robust noisy label domain adaptation method based on confidence and information entropy, characterized by: The specific steps include: S1. Prepare data in the source domain and target domain dual-domain datasets; S2, the warm-up phase, where the model is trained using labeled data from the source domain; S3, confidence modulation adaptive learning; S4, entropy-guided obfuscation alignment mechanism; S5. Establish a joint optimization loss function; S6. Output the adaptive result.

2. The method for robust noise label domain adaptation based on confidence and information entropy according to claim 1, characterized in that: The specific method of step S1 is as follows: Dual domain dataset D = {I s , I t }Including source domain data I s and target domain data I t , the source domain data contains N s Sample pairs represents the i-th source domain sample, Represents the corresponding category label, the category label value range is {1, 2, ..., C}, the target domain data contains N t samples The original data is input into the deep convolutional neural network model, and the input image data is normalized, its pixel values are scaled to the range of [0, 1], and standardized according to the ImageNet pixel value mean μ = [0.485, 0.456, 0.406] and pixel value standard deviation σ = [0.229, 0.224, 0.225]. The expression is as follows: Among them, x norm represents the normalized image data, and x represents the pixel value of the original image data; A certain proportion of symmetric noise and asymmetric noise are injected into the source domain data, and ResNet-50 is used as the basic backbone network to test the robustness under different noise intensities.

3. The method for robust noise label domain adaptation based on confidence and information entropy according to claim 1, characterized in that: The specific method of step S2 is as follows: After data preparation is completed, the early learning characteristics of deep neural networks are used to quickly establish feature representation in the initial stage. The model uses source domain data samples to Optimize the loss function, the specific expression is as follows: in, represents the classification loss of source domain samples, N s represents the total number of source domain samples; The warm-up phase lasts for T warm During the training process, the model performance is evaluated on the source domain data subset every several steps. After the warm-up is completed, the model parameters are saved to provide a good initialization for the subsequent training stages.

4. The method for robust noise label domain adaptation based on confidence and information entropy according to claim 1, characterized in that: The specific method of step S3 is as follows: Through the domain shared network f = [f1, ..., f C ] to predict the dual domain data, where f i Represents the output of the i-th category, and the network prediction result is Contains source domain prediction probability and target domain prediction probability The predicted probability is calculated by the softmax function, and the specific expression is as follows: in, Represents source domain samples The network output score on the jth class, Represents source domain samples The network output score on the kth class, Represents the target domain sample The network output score on the jth class, Represents the target domain sample The network output score on the kth class, τ represents the temperature parameter; Using the model parameters and feature representations obtained in the warm-up phase, we introduce a mechanism to dynamically adjust the learning intensity of samples based on prediction confidence and construct an adaptive loss function based on prediction confidence. The expression is as follows: in, represents the adaptive loss of a single sample, represents the predicted probability of the i-th source domain sample for the j-th class, and q represents the adaptive modulation parameter; The overall loss is calculated as follows: Here, N represents the batch size.

5. The method for robust noise label domain adaptation based on confidence and information entropy according to claim 1, characterized in that: The specific method of step S4 is as follows: For the target domain sample, the target domain prediction probability is used to calculate its entropy value. The specific expression is as follows: in, Represents the target domain sample The predicted probability of category j, C represents the total number of categories; Based on the entropy value of the target sample, an entropy-guided weight generation function is designed to capture the uncertainty of the sample. The specific expression is as follows: Among them, α represents the parameter that controls the steepness of the weight distribution, e i Represents the target domain sample The entropy value of In order to capture the relationship between categories, the predicted probability of the target sample is used to generate the category alignment matrix, which is expressed as follows: Among them, diag(w) means according to w i The resulting diagonal matrix, A ij represents the element in the i-th row and j-th column of matrix A; By performing row normalization on the category alignment matrix A, the normalized category alignment matrix is obtained The expression is as follows: In order to reduce the uncertainty between categories, a category confusion loss function based on entropy is designed. The specific expression is as follows: in, Representation matrix traces.

6. The method for robust noise label domain adaptation based on confidence and information entropy according to claim 1, characterized in that: The specific method of step S5 is as follows: Through the confidence-controlled adaptive learning loss and the entropy-guided confusion alignment loss, robustness optimization in cross-domain adaptation from the source domain to the target domain is achieved, and the expression of the joint optimization loss function is obtained as follows: THE total =L CMAL +λL EGCA Where λ represents the balance parameter, which is used to adjust L CMAL and L EGCA The relative importance of joint optimization.

7. The method for robust noise label domain adaptation based on confidence and information entropy according to claim 1, characterized in that: The specific method of step S6 is as follows: The classification accuracy is used to evaluate the overall prediction accuracy of the model on the target domain samples. The classification accuracy is expressed as follows: Among them, N t represents the total number of target domain samples, y t represents the true label of the target domain sample, represents the indicator function, Represents the predicted label of the target domain sample; if the prediction result Equal to the true label y t , the indicator function takes the value 1, otherwise it is 0.

Images