Noise label robust learning method and system based on DI and SNOP
By employing a robust learning method for noisy labels based on DI and SNOP, and utilizing dissimilar signals and orthogonal feature constraints, the robustness problem of deep learning models under noisy labels is solved, achieving significant performance improvement in multimodal scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING INST OF TECH
- Filing Date
- 2025-12-01
- Publication Date
- 2026-05-08
AI Technical Summary
Existing technologies struggle to improve the robustness and generalization ability of deep learning models in noisy label environments. In particular, the presence of noisy labels can easily disrupt semantic similarity, leading to incorrect associations learned by the model and impacting performance.
We adopt a robust learning method for noise labels based on DI and SNOP, using dissimilar signals as robust anchors. We combine structured orthogonal constraints in the feature space with adaptive similarity adjustment, and construct a robust deep learning framework through cross-entropy loss, structured negative orthogonal penalty loss, and dissimilarity calibration similarity adjustment loss.
It significantly improves the robustness and generalization ability of deep learning models in high-noise environments, prevents false similarity caused by noisy labels, and is suitable for classification tasks in multimodal scenarios.
Smart Images
Figure CN121997109A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a robust learning method and system for noisy labels based on DI and SNOP, belonging to the field of deep learning model training and optimization technology, and is applicable to multimodal classification tasks such as image classification, speech recognition, and natural language processing when the labeled data contains noise or mislabeling. Background Technology
[0002] Deep learning technology has made significant progress in computer vision, speech recognition, and natural language processing in recent years. However, in practical applications, training data often struggles to guarantee complete label accuracy, leading to the problem of noisy labels. Existing methods for learning noisy labels mainly employ the following approaches: 1. Loss correction methods: These methods mitigate the impact of noisy labels by estimating the noise transition matrix or modifying the loss function. However, the noise transition matrix is difficult to estimate accurately, and uniformly correcting all samples can lead to accumulated errors, affecting model performance. 2. Sample selection methods: These methods select "clean" samples for training based on loss values or feature similarity, discarding potentially noisy samples. However, this method typically discards a large amount of data, resulting in knowledge waste. 3. Semi-supervised learning methods: These methods treat potentially noisy samples as unlabeled data, using semi-supervised learning strategies to improve the model's robustness. This method is highly sensitive to hyperparameter settings and increases training complexity. These methods largely rely on sample similarity as a learning signal, but in the presence of noisy labels, semantic similarity is easily disrupted, causing the model to learn incorrect associations and thus reducing generalization performance.
[0003] However, in noisy labeling environments, semantic dissimilarity is termed dissimilarity invariance (DI). Compared to fragile similarity signals, dissimilarity signals are less sensitive to noise and can serve as more reliable training anchors. However, current technologies have not systematically applied this phenomenon to noisy labeling learning frameworks, nor do they possess a systematic approach to combining it with feature orthogonality constraints and adaptive similarity adjustment.
[0004] Therefore, improving the robustness of deep learning models in noisy labeling scenarios for classification tasks has become an urgent problem to be solved. Summary of the Invention
[0005] The purpose of this invention is to address the technical problem of improving the robustness of deep learning models in noisy label scenarios during classification tasks. It proposes a robust learning method and system based on DI and SNOP for noisy labels. By utilizing dissimilarity signals that remain stable in noisy label environments as robust anchors, and combining structured orthogonal constraints in the feature space with adaptive similarity adjustment, it effectively suppresses spurious similarities caused by noisy labels, thereby significantly improving the model's robustness and generalization ability in high-noise environments.
[0006] The objective of this invention is achieved through the following technical solution:
[0007] This invention discloses a robust learning method for noise labels based on DI and SNOP, comprising the following:
[0008] step:
[0009] Step 1: Construct a ResNet-18 neural network model and randomly initialize the model parameters θ;
[0010] Step 2: Use a feature extractor to extract features from the training dataset containing noisy labels to obtain feature vectors;
[0011] Step 2.1: Obtain the training dataset containing noisy labels as shown in Equation (1);
[0012]
[0013] Where, x i Indicates input data, For potentially contaminated labels, N is the total number of samples;
[0014] Step 2.2: Use the feature extractor f(·θ) to extract features from the training dataset and obtain feature vectors.
[0015] Step 3: Divide the set of semantically related sample pairs and the set of semantically unrelated sample pairs according to the semantic distance of the label pairs based on the set threshold;
[0016] Step 3.1: Use the class hierarchy structure of Dissimilarity Invariance (DI) to obtain the semantic distance of the label pairs consisting of the sample labels of the two training datasets as shown in Equation (2);
[0017]
[0018] Where h(·) is the depth function of the category hierarchy tree, and NL(·) represents the lowest common ancestor node;
[0019] Step 3.2: Set the semantic dissimilarity threshold η min ,η max According to the set threshold, the sample pairs are divided into a set of semantically related samples and a set of semantically unrelated samples as shown in Equation (3);
[0020]
[0021] Among them, the semantically related sample set is Semantically irrelevant sample set
[0022] Step 4: Utilize cross-entropy loss, structured negative orthogonal penalty loss, and dissimilarity calibration similarity adjustment loss to obtain the total loss of the ResNet-18 neural network model;
[0023] Step 4.1: Obtain the structured negative orthogonal penalty loss for semantically unrelated samples;
[0024] Step 4.1.1: Apply the Structured Negative Orthogonality Penalty (SNOP) method to the semantically irrelevant sample set. Construct the difference vector set;
[0025]
[0026] Step 4.1.2: Use equation (5) to obtain the global orthogonality loss and the local confidence-weighted orthogonality loss respectively;
[0027]
[0028] Among them, w ij =(1-c i )·(1-c j );c i Let i be the prediction confidence of sample i, that is, the softmax probability of its predicted class;
[0029] Step 4.1.3: Use global orthogonality loss and local confidence weighted orthogonality loss to obtain the SNOP loss of the difference vector set as shown in Equation (6);
[0030] L SNOP =L global +L local (6)
[0031] Step 4.2: Obtain the dissimilarity calibration similarity adjustment loss for semantically related samples;
[0032] Step 4.2.1: Dissimilarity-Calibrated Similarity Adjustment (DCSA) is applied to semantically relevant sample sets. Obtain the calibration factor as shown in equation (7);
[0033]
[0034] Where p and q represent dissimilarity anchors, i.e., the corresponding semantically irrelevant samples; i and j represent the corresponding semantically relevant samples. Transpose of a vector;
[0035] Step 4.2.2: Use cosine similarity to obtain the DCSA loss as shown in equation (8);
[0036]
[0037] in Cosine similarity;
[0038] Step 4.3: Using cross-entropy loss, structured negative orthogonal penalty loss and dissimilarity calibration similarity adjustment loss, the total loss of the ResNet-18 neural network model as shown in Equation (9) is obtained;
[0039] L total =L org +λL SNOP +μL DCSA (9)
[0040] Among them, L org The cross-entropy loss is λ, and the balance coefficients are μ.
[0041] Step 5: Update the model parameters θ using the total loss of the ResNet-18 neural network model through backpropagation to obtain the trained ResNet-18 neural network model;
[0042] Step 6: Input the test data into the trained ResNet-18 neural network model to obtain the results of the multimodal scene classification task;
[0043] To achieve the objectives of this invention, based on the above method, this invention further proposes a robust learning system for noise labels based on DI and SNOP, including a data processing module, a feature extraction module, a dissimilarity analysis module, a SNOP module, a DCSA module, and a loss calculation and model training module.
[0044] The data processing module is used to read the training dataset containing noisy labels and perform data cleaning and preprocessing; this data will then serve as input to the feature extraction module.
[0045] The feature extraction module maps input samples to a feature representation space, performs feature extraction on the dataset, and outputs the corresponding feature vectors as results; these vectors will serve as inputs to the dissimilarity analysis module, the SNOP module, and the DCSA module.
[0046] The dissimilarity analysis module is used to divide sample pairs into semantically related sets based on the semantic distance obtained from the class hierarchy and according to a set threshold. Set of semantically unrelated sample pairs It will be used as input to the SNOP module and the DCSA module;
[0047] The SNOP module is used to obtain a set of feature difference vectors for semantically unrelated sample pairs and apply global and local orthogonality constraints to obtain a structured negative orthogonality penalty loss, which will be used as input to the loss calculation and model training modules.
[0048] The DCSA module is used to obtain calibration factors for semantically related sample pairs through dissimilarity anchors and set an adaptive soft upper limit on similarity. Positive sample pairs that exceed the upper limit are penalized to obtain dissimilarity calibration similarity adjustment loss, which will be used as input to the loss calculation and model training modules.
[0049] The loss calculation and model training module is used to calculate the cross-entropy loss of the model on the training dataset, and to combine the cross-entropy loss, SNOP loss and DCSA loss into a weighted total loss function. The module also uses optimization algorithms to iteratively update the model parameters to achieve robust end-to-end training.
[0050] Beneficial effects:
[0051] Compared with existing technologies, it has the following beneficial effects:
[0052] 1. Utilize the invariance of dissimilarity as a stable training anchor point to avoid relying on similarity signals that are susceptible to noise interference.
[0053] 2. By using structured negative orthogonal penalty (SNOP), the effective separation of unrelated sample pairs is maintained in the feature space, reducing the risk of propagation of erroneous associations.
[0054] 3. By using Dissimilarity Calibration Similarity Adjustment (DCSA), we can dynamically suppress the spurious similarity boost caused by noise labels and prevent the model from overfitting incorrect relationships.
[0055] 4. The method is designed as a plug-in, which can be seamlessly integrated into existing deep learning frameworks without the need for additional external data or pre-training, and has a wide range of applications.
[0056] 5. Significant robustness improvements have been verified on various synthetic and real noise datasets, demonstrating promising prospects for industrial applications. Attached Figure Description
[0057] Figure 1 This is a flowchart illustrating the present invention; Detailed Implementation
[0058] To better illustrate the purpose and advantages of the present invention, the invention will be further described below with reference to the accompanying drawings and examples. It should be noted that the implementation of the present invention is not limited to the following embodiments, and any modifications or alterations made to the present invention will fall within the protection scope of the present invention.
[0059] Example
[0060] like Figure 1 As shown in the figure, the specific implementation steps of a robust learning method for noise labels based on DI and SNOP in this embodiment are as follows:
[0061] Step 1: Construct a ResNet-18 neural network model and randomly initialize the model parameters θ;
[0062] Step 2: Use a feature extractor to extract features from the training dataset containing noisy labels to obtain feature vectors;
[0063] Step 2.1: Obtain the training dataset containing noisy labels as shown in Equation (1);
[0064]
[0065] Where, x i Indicates input data, For potentially contaminated labels, N is the total number of samples;
[0066] Step 2.2: Use the feature extractor f(·θ) to extract features from the training dataset and obtain feature vectors.
[0067] In this embodiment, in terms of data input and feature extraction, a training dataset containing noisy labels is obtained, and the samples are input into a feature extractor to obtain feature representation vectors.
[0068] Step 3: Divide the set of semantically related sample pairs and the set of semantically unrelated sample pairs according to the semantic distance of the label pairs based on the set threshold;
[0069] Step 3.1: Use the class hierarchy structure of Dissimilarity Invariance (DI) to obtain the semantic distance of the label pairs consisting of the sample labels of the two training datasets as shown in Equation (2);
[0070]
[0071] Where h(·) is the depth function of the category hierarchy tree, and NL(·) represents the lowest common ancestor node;
[0072] Step 3.2: Set the semantic dissimilarity threshold η min ,η max According to the set threshold, the sample pairs are divided into a set of semantically related samples and a set of semantically unrelated samples as shown in Equation (3);
[0073]
[0074] Among them, the semantically related sample set is Semantically irrelevant sample set
[0075] In the embodiments, the semantic distance of sample pairs is calculated by class taxonomy in the modeling of dissimilarity invariance, distinguishing between semantically related sample pairs and semantically unrelated sample pairs, and a robust set of dissimilarity anchor points is constructed by utilizing the stability of dissimilarity under different noise ratios.
[0076] Step 4: Utilize cross-entropy loss, structured negative orthogonal penalty loss, and dissimilarity calibration similarity adjustment loss to obtain the total loss of the ResNet-18 neural network model;
[0077] Step 4.1: Obtain the structured negative orthogonal penalty loss for semantically unrelated samples;
[0078] Step 4.1.1: Apply the Structured Negative Orthogonality Penalty (SNOP) method to the semantically irrelevant sample set. Construct the difference vector set;
[0079]
[0080] Step 4.1.2: Use equation (5) to obtain the global orthogonality loss and the local confidence-weighted orthogonality loss respectively;
[0081]
[0082] Among them, w ij =(1-c i )·(1-c j );c i Let i be the prediction confidence of sample i, that is, the softmax probability of its predicted class;
[0083] Step 4.1.3: Use global orthogonality loss and local confidence weighted orthogonality loss to obtain the SNOP loss of the difference vector set as shown in Equation (6);
[0084] L SNOP =L global +L local (6)
[0085] In the embodiment, for the structured negative orthogonal penalty (SNOP), the set of feature difference vectors is calculated for irrelevant sample pairs, and global orthogonality constraints and local confidence-weighted orthogonality constraints are applied: the global orthogonality constraint minimizes the difference between the difference vector Gram matrix and the identity matrix, so that the negative sample pairs are orthogonally distributed globally; the local constraint applies a stronger orthogonal penalty to low-confidence sample pairs based on the predicted confidence of the samples.
[0086] Step 4.2: Obtain the dissimilarity calibration similarity adjustment loss for semantically related samples;
[0087] Step 4.2.1: Dissimilarity-Calibrated Similarity Adjustment (DCSA) is applied to semantically relevant sample sets. Obtain the calibration factor as shown in equation (7);
[0088]
[0089] Where p and q represent dissimilarity anchors, i.e., the corresponding semantically irrelevant samples; i and j represent the corresponding semantically relevant samples. Transpose of a vector;
[0090] Step 4.2.2: Use cosine similarity to obtain the DCSA loss as shown in equation (8);
[0091]
[0092] in Cosine similarity;
[0093] In the embodiment, for the Dissimilarity Calibration Similarity Adjustment (DCSA), for semantically related sample pairs, the degree of proximity between them and unrelated samples is calculated as a calibration factor. An adaptive soft upper limit is set for the similarity of positive sample pairs based on this factor, and similarity values exceeding the upper limit are penalized to prevent erroneous high similarity caused by noise labels.
[0094] Step 4.3: Using cross-entropy loss, structured negative orthogonal penalty loss and dissimilarity calibration similarity adjustment loss, the total loss of the ResNet-18 neural network model as shown in Equation (9) is obtained;
[0095] L total =L org +λL SNOP +μL DCSA (9)
[0096] Among them, L org The cross-entropy loss is λ, and the balance coefficients are μ.
[0097] Step 5: Update the model parameters θ using the total loss of the ResNet-18 neural network model through backpropagation to obtain the trained ResNet-18 neural network model;
[0098] In this embodiment, in terms of loss function construction and model update, the original task loss is combined with SNOP loss and DCSA loss in a weighted manner to obtain the total loss function, and the model parameters are updated through backpropagation to achieve robust feature learning.
[0099] Step 6: Input the test data into the trained ResNet-18 neural network model to obtain the results of the multimodal scene classification task;
[0100] To achieve the objectives of this invention, based on the above method, this embodiment further proposes a robust learning system for noise labels based on DI and SNOP, including a data processing module, a feature extraction module, a dissimilarity analysis module, a SNOP module, a DCSA module, and a loss calculation and model training module;
[0101] The data processing module is used to read the training dataset containing noisy labels and perform data cleaning and preprocessing; this data will then serve as input to the feature extraction module.
[0102] The feature extraction module maps input samples to a feature representation space, performs feature extraction on the dataset, and outputs the corresponding feature vectors as results; these vectors will serve as inputs to the dissimilarity analysis module, the SNOP module, and the DCSA module.
[0103] The dissimilarity analysis module is used to divide sample pairs into semantically related sets based on the semantic distance obtained from the class hierarchy and according to a set threshold. Set of semantically unrelated sample pairs It will be used as input to the SNOP module and the DCSA module;
[0104] The SNOP module is used to obtain a set of feature difference vectors for semantically unrelated sample pairs and apply global and local orthogonality constraints to obtain a structured negative orthogonality penalty loss, which will be used as input to the loss calculation and model training modules.
[0105] The DCSA module is used to obtain calibration factors for semantically related sample pairs through dissimilarity anchors and set an adaptive soft upper limit on similarity. Positive sample pairs that exceed the upper limit are penalized to obtain dissimilarity calibration similarity adjustment loss, which will be used as input to the loss calculation and model training modules.
[0106] The loss calculation and model training module is used to calculate the cross-entropy loss of the model on the training dataset, and to combine the cross-entropy loss, SNOP loss and DCSA loss into a weighted total loss function. The module also uses optimization algorithms to iteratively update the model parameters to achieve robust end-to-end training.
[0107] To further illustrate the advantages of the present invention, it will be described in conjunction with application scenarios.
[0108] 1. Image classification scenario
[0109] In large-scale image crawling tasks on the internet, the collected images typically rely on crawlers to automatically extract category information from webpage text, tags, or metadata. These tags often contain a high percentage of errors. For example, some cat photos are labeled as dogs. This invention ensures sufficient separation of feature differences between different categories by using the stability of dissimilarity anchors and the global orthogonality constraint of the SNOP module. Simultaneously, the DCSA module prevents the model from learning false high similarity caused by noisy labels, thus significantly improving classification accuracy even with a noise ratio as high as 40%.
[0110] 2. Medical imaging diagnostic scenarios
[0111] In medical image annotation, differences in annotators' experience, visual judgment biases, and inconsistencies in clinical standards can lead to errors in the labeling of images such as X-rays, CT scans, and MRIs. For example, a lung CT scan showing a minor lesion might be mislabeled as normal. This invention utilizes the invariance of dissimilarity to prioritize maintaining stable feature differences between different disease categories during training, while suppressing erroneous cross-class high similarity caused by noisy labels. This effectively improves the robustness and stability of the model in medical image diagnosis tasks and reduces the misdiagnosis rate.
[0112] 3. Autonomous driving perception scenarios
[0113] In autonomous driving perception systems, target detection and classification rely on the fusion of multiple sensor sources (such as cameras, LiDAR, and millimeter-wave radar). During the fusion process, labeling errors may occur due to time synchronization errors, sensor errors, or labeling mistakes. For example, a pedestrian may be mislabeled as a cyclist. The SNOP module of this invention ensures that the features of different categories such as vehicles, pedestrians, and cyclists are orthogonally distributed in a high-dimensional space, avoiding feature confusion; the DCSA module prevents erroneous clustering effects caused by a small number of mislabeled features, thereby ensuring stable perception and decision-making of the autonomous driving system in complex road environments.
[0114] 4. Speech and text recognition scenarios
[0115] In speech recognition and text recognition tasks, the results of manual transcription and automatic speech recognition may contain errors such as misspellings, omissions, and homophone confusion. For example, "bank" might be incorrectly identified as "welcoming." This invention utilizes dissimilarity anchors and a similarity adaptive adjustment mechanism to ensure that features of different semantic categories remain stably distinguishable during training, preventing the cumulative effect of erroneous similarity caused by noise labels, thereby improving the generalization ability and robustness of speech and text recognition models.
[0116] 5. Cross-modal retrieval scenarios
[0117] In cross-modal retrieval tasks (such as image-text retrieval and video-text retrieval), training data typically comes from the internet, and its tag correspondences (such as matching images with descriptive text) may contain mismatches. This invention can stably maintain the separation of different semantic categories in the cross-modal space through a dissimilarity invariance mechanism, thereby reducing the performance degradation caused by mismatches.
[0118] Table 1 Comparison of Results in CIFAR-N
[0119]
[0120] The results, as shown in Table 1, demonstrate that the combination of this invention and RoLR can significantly improve robustness when performing image classification tasks in real-world noisy environments.
Claims
1. A robust learning method for noise labels based on DI and SNOP, characterized in that: Includes the following steps, Step 1: Construct a ResNet-18 neural network model and randomly initialize the model parameters θ; Step 2: Use a feature extractor to extract features from the training dataset containing noisy labels to obtain feature vectors; Step 3: Divide the set of semantically related sample pairs and the set of semantically unrelated sample pairs according to the semantic distance of the label pairs based on the set threshold; Step 3.1: Utilize the class hierarchy structure with dissimilarity invariance to obtain the semantic distance of the label pairs consisting of sample labels from the two training datasets, as shown in Equation (2); Where h(·) is the depth function of the category hierarchy tree, and NL(·) represents the lowest common ancestor node; Step 3.2: Set the semantic dissimilarity threshold η min ,η max According to the set threshold, the sample pairs are divided into a set of semantically related samples and a set of semantically unrelated samples as shown in Equation (3); Among them, the semantically related sample set is Semantically irrelevant sample set Step 4: Utilize cross-entropy loss, structured negative orthogonal penalty loss, and dissimilarity calibration similarity adjustment loss to obtain the total loss of the ResNet-18 neural network model; Step 5: Update the model parameters θ using the total loss of the ResNet-18 neural network model through backpropagation to obtain the trained ResNet-18 neural network model; Step 6: Input the test data into the trained ResNet-18 neural network model to obtain the results of the multimodal scene classification task.
2. The robust learning method for noise labels based on DI and SNOP as described in claim 1, characterized in that: Step 2 is implemented as follows: Step 2.1: Obtain the training dataset containing noisy labels as shown in Equation (1); Where, x i Indicates input data, For potentially contaminated labels, N is the total number of samples; Step 2.2: Use the feature extractor f(·θ) to extract features from the training dataset and obtain feature vectors.
3. The robust learning method for noise labels based on DI and SNOP as described in claim 1, characterized in that: Step 4 is implemented as follows: Step 4.1: Obtain the structured negative orthogonal penalty loss for semantically unrelated samples; Step 4.2: Obtain the dissimilarity calibration similarity adjustment loss for semantically related samples; Step 4.3: Using cross-entropy loss, structured negative orthogonal penalty loss and dissimilarity calibration similarity adjustment loss, the total loss of the ResNet-18 neural network model as shown in Equation (9) is obtained; L total =L org +λL SNOP +μL DCSA (9) Among them, L org λ represents the cross-entropy loss, and μ represents the balance coefficients.
4. The robust learning method for noise labels based on DI and SNOP as described in claim 3, characterized in that: Step 4.1 is implemented as follows: Step 4.1.1: Apply the Structured Negative Orthogonality Penalty (SNOP) method to the semantically irrelevant sample set. Construct the difference vector set; Step 4.1.2: Use equation (5) to obtain the global orthogonality loss and the local confidence-weighted orthogonality loss respectively; Among them, w ij =(1-c i )·(1-c j );c i Let i be the prediction confidence of sample i, that is, the softmax probability of its predicted class; Step 4.1.3: Use global orthogonality loss and local confidence weighted orthogonality loss to obtain the SNOP loss of the difference vector set as shown in Equation (6). L SNOP =L global +L local (6) 5. The robust learning method for noise labels based on DI and SNOP as described in claim 3, characterized in that: Step 4.2 is implemented as follows: Step 4.2.1: Dissimilarity-Calibrated Similarity Adjustment (DCSA) is applied to semantically related sample sets. Obtain the calibration factor as shown in equation (7); Where p and q represent dissimilarity anchors, i.e., the corresponding semantically irrelevant samples; i and j represent the corresponding semantically relevant samples; and T is the vector transpose. Step 4.2.2: Use cosine similarity to obtain the DCSA loss as shown in equation (8); in Let be the cosine similarity.
6. A robust noise label learning system based on DI and SNOP, implementing the method as described in claim 1, characterized in that: It includes a data processing module, a feature extraction module, a dissimilarity analysis module, a SNOP module, a DCSA module, and a loss calculation and model training module; The data processing module is used to read the training dataset containing noisy labels and perform data cleaning and preprocessing; this data will then serve as input to the feature extraction module. The feature extraction module is used to map input samples to a feature representation space, perform feature extraction operations on the dataset, and output the corresponding feature vector as the result. It will serve as input for the dissimilarity analysis module, the SNOP module, and the DCSA module; The dissimilarity analysis module is used to divide sample pairs into semantically related sets based on the semantic distance obtained from the class hierarchy and according to a set threshold. Set of semantically unrelated sample pairs It will be used as input to the SNOP module and the DCSA module; The SNOP module is used to obtain a set of feature difference vectors for semantically unrelated sample pairs and apply global and local orthogonality constraints to obtain a structured negative orthogonality penalty loss, which will be used as input to the loss calculation and model training modules. The DCSA module is used to obtain calibration factors for semantically related sample pairs through dissimilarity anchors and set an adaptive soft upper limit on similarity. Positive sample pairs that exceed the upper limit are penalized to obtain dissimilarity calibration similarity adjustment loss, which will be used as input to the loss calculation and model training modules. The loss calculation and model training module is used to calculate the cross-entropy loss of the model on the training dataset, and to combine the cross-entropy loss, SNOP loss and DCSA loss into a weighted total loss function. The module also uses optimization algorithms to iteratively update the model parameters to achieve robust end-to-end training.