Offset mark learning representation enhancement method and system based on representation difference alignment
By adopting a partial label learning method based on representation difference alignment, the problem of inconsistent differences between categories in neural network models in image classification tasks is solved, which improves the representation quality and classification accuracy of the model and achieves stronger robustness and generalization ability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTHEAST UNIV
- Filing Date
- 2026-01-13
- Publication Date
- 2026-05-12
AI Technical Summary
Existing partial label learning methods, such as ResNet and ConvNet, in image classification tasks suffer from inconsistent differences between categories within the representation space, affecting classification generalization performance.
We employ a partial label learning method based on representation difference alignment, using KL divergence to estimate inter-class relationships, regularize and reduce the dimensionality of the representation space, and combine a random representation dimension masking strategy to enhance the global semantic topology of the model.
It improves the model's representation quality and classification performance, increases the accuracy of image classification, enhances the method's robustness and generalization ability, and is plug-and-play.
Smart Images
Figure CN122024006A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence and machine learning, and specifically relates to a method and system for enhancing representations by partial label learning based on representation difference alignment. Background Technology
[0002] Partially labeled learning is an important paradigm in weakly supervised learning, in which each training instance is labeled with a set of candidate labels, which contains only one correct label but the specific label is unknown. The inherent ambiguity of labels poses a fundamental challenge to learning reliable decision boundaries and effective generalization.
[0003] Existing partial label learning methods can be broadly categorized into two types: averaging-based methods treat candidate labels equally to construct a supervisory signal, but are severely affected by label noise due to their inability to distinguish between correct and incorrect candidates; recognition-based methods explicitly identify the true label for each instance during training, and the most advanced methods in recent years mainly belong to this category. Although these methods have achieved success on controlled benchmarks, they often perform poorly in more realistic scenarios (such as instance-related candidate sets).
[0004] A key observation from recent advances is that high-quality representation learning plays a crucial role in resolving label ambiguity. Models are more likely to infer correct labels based on neighborhood structure if semantically similar instances cluster together in the representation space and different classes are well separated. This insight has inspired a range of representation-driven partial label learning methods. However, these methods primarily focus on local objectives (such as minimizing intra-class variance or maximizing inter-class separation) without explicitly modeling the global structure of inter-class relationships in the representation space. While this local optimization strategy is beneficial, it often leads to imbalanced class separation in the representation space: different class pairs may exhibit significantly different degrees of separation, with some being excessively pushed apart while others remain relatively close. This structural imbalance can impair the semantic consistency of learned representations, thereby affecting generalization performance. Summary of the Invention
[0005] Purpose of the invention: This invention addresses the problem that existing partial label learning methods in image classification tasks suffer from inconsistent differences between categories in the representation space, affecting classification generalization performance. It proposes a partial label learning representation enhancement method and system based on representation difference alignment, enabling the partial label learning model to learn a globally balanced semantic topology, thereby improving the model's representation quality and classification performance.
[0006] Technical solution: The present invention provides a method for enhancing representations based on representation difference alignment using partial label learning, comprising the following steps:
[0007] (1) Based on a partially labeled dataset consisting of multiple image samples and its candidate label set, the visual class relationship between different image samples is estimated in a label-independent manner using a label distribution difference measure based on KL divergence.
[0008] (2) Based on the visual category relationships estimated in step (1), the visual category representation differences in the high-dimensional image representation space output by the neural network image encoder are aligned and regularized.
[0009] (3) Combining the random representation dimension masking strategy, the high-dimensional image representation vector is sampled in a way that reduces the computational complexity of image representation difference alignment;
[0010] (4) The representation difference alignment framework is seamlessly integrated with the existing partial label learning image classification method based on deep neural networks, and the model parameters used for image feature extraction are iteratively updated to improve the classification accuracy of the image classification model for unknown image samples.
[0011] Furthermore, the implementation process of step (1) is as follows:
[0012] The label distribution difference measure based on KL divergence is defined as:
[0013]
[0014] in, For the sample and The difference in label distribution between them is a metric. Indicates the first Samples during the next training iteration The tags are distributed across categories The value on, Indicates the first Samples during the next training iteration The tags are distributed across categories The value on, For adjustable hyperparameters, Total number of categories;
[0015] Using confidence threshold Filtering low-confidence samples, the weights of sample pairs in the difference alignment loss are defined as follows:
[0016] .
[0017] Furthermore, the alignment regularization of the representation differences between visual categories adopts a relative difference form, and the regularization loss function is defined as:
[0018]
[0019] in, For first-order representation differences, For second-order representation differences, For encoder parameters, For the sample size, For sample pairs and The final weights in the difference alignment loss, For sample pairs and The final weights in the difference alignment loss.
[0020] Furthermore, the first-order representation difference is the Euclidean distance between the representations of the two samples:
[0021]
[0022] in, For encoder, For encoder parameters, and There are two input samples;
[0023] The difference in second-order representations is the absolute difference between first-order differences:
[0024] .
[0025] Furthermore, step (3) is implemented as follows:
[0026] Random representation dimension masking strategy uses mask vectors Implementation, in which , For the original representation dimension, To reduce the dimensionality, the difference in first-order representations after masking is redefined as:
[0027]
[0028] in, This indicates element-wise multiplication.
[0029] Furthermore, the implementation process of step (4) is as follows:
[0030] The representations of all supervised views are integrated into a unified representation space, in which inter-category difference alignment is performed;
[0031] The final total loss function is a weighted sum of the original loss from partial label learning and the loss from aligning representation differences:
[0032]
[0033] in, The original loss function for the selected partial label learning method. To balance the hyperparameters of the two losses, This is the regularization loss function.
[0034] The present invention discloses a partial label learning representation enhancement system based on representation difference alignment, comprising:
[0035] The relationship estimation module takes a partially labeled dataset consisting of multiple image samples and its candidate label set as input, and uses a label distribution difference measure based on KL divergence to estimate the visual category relationship between different image samples in a label-independent manner.
[0036] The difference alignment module aligns and regularizes the differences in visual category representations in the high-dimensional image representation space output by the neural network image encoder based on the relationship between visual categories.
[0037] The dimension masking module, combined with a random representation dimension masking strategy, performs dimensionality reduction sampling on high-dimensional image representation vectors to reduce the computational complexity of image representation difference alignment.
[0038] The model integration module seamlessly integrates the representation difference alignment framework with existing deep neural network-based partial label learning image classification methods, and iteratively updates the model parameters used for image feature extraction.
[0039] The present invention provides a storage medium storing a computer program, which, when executed by at least one processor, implements the steps of the partial label learning representation enhancement method based on representation difference alignment as described above.
[0040] An electronic device according to the present invention includes a memory and a processor, wherein:
[0041] Memory is used to store computer programs that can run on a processor;
[0042] A processor, configured to, while running the computer program, execute the steps of the partial label learning representation enhancement method based on representation difference alignment as described above.
[0043] Beneficial Effects: Compared with existing technologies, the beneficial effects of this invention are as follows: 1. This invention, through a representation difference alignment regularization framework, explicitly constrains the consistency of differences between categories in the representation space, solving the problem of existing partial label learning methods neglecting global structure, and improving the structural quality of representations and the generalization ability of models; 2. The label distribution difference metric based on KL divergence proposed in this invention can reliably estimate the relationship between categories in a label-independent manner in the early training stage of partial label learning, enhancing the robustness of the method; 3. The random representation dimension masking strategy introduced in this invention effectively reduces the computational complexity of high-dimensional representation difference alignment while maintaining global structural information; 4. The framework of this invention is fully compatible with existing partial label learning methods, can be seamlessly integrated without changing the original architecture, and has plug-and-play convenience. Attached Figure Description
[0044] Figure 1 Flowchart of a partial label learning representation enhancement method based on representation difference alignment;
[0045] Figure 2 A diagram of the architecture of a partial label learning representation enhancement model for aligning representation differences. Detailed Implementation
[0046] The present invention will now be described in further detail with reference to the accompanying drawings.
[0047] This invention proposes a partial label learning representation enhancement method based on representation difference alignment to address the technical problem of insufficient model generalization performance caused by the imbalance of inter-class representation differences in image classification tasks. The method mainly includes: using a label distribution difference metric based on KL divergence to estimate the sample relationship between classes in a label-independent manner; performing alignment regularization on the inter-class representation differences in the representation space based on the estimated inter-class relationship; combining a random representation dimension masking strategy to reduce the computational complexity of high-dimensional representation difference alignment; and seamlessly integrating the representation difference alignment framework with existing partial label learning methods to iteratively update the model parameters.
[0048] Existing partial labeling learning methods typically focus on local discriminative objectives while neglecting the global structure of the representation space, leading to imbalanced class separation and impaired generalization performance. This embodiment promotes globally consistent inter-class relationships in the representation space through a representational difference alignment framework, complementing existing local discriminative objectives. Figure 1 and Figure 2 As shown, the specific steps include the following:
[0049] S1: Inter-class relationship estimation stage: Input samples and their candidate label sets from the partially labeled dataset, and estimate the degree of class difference between sample pairs using a label distribution difference measure based on KL divergence.
[0050] Specifically, in the early training phase of partial label learning, reliable label information is often insufficient to determine whether two samples belong to different categories. To address this issue, a label distribution dissimilarity metric based on KL divergence is used to quantify the dissimilarity of label distributions:
[0051]
[0052] in, For the sample and The difference in label distribution between them is a metric. Characterizing the first Samples during the next training iteration Label distribution Indicates the first Samples during the next training iteration The tags are distributed across categories The value on, For adjustable hyperparameters, This represents the total number of categories.
[0053] Considering the low reliability of early label disambiguation results in partial label learning, a confidence threshold is used. Filtering low-confidence samples, the final weight of sample pairs in the difference alignment loss is defined as follows:
[0054]
[0055] This formula adaptively quantifies label differences by utilizing the divergence between probabilistic predictions, avoiding reliance on unreliable pseudo-labels in the early training phase.
[0056] S2: Representation Difference Alignment Stage: Based on the inter-class relationship weights estimated in step S1, the first-order and second-order representation differences in the representation space are calculated, and a relative form of representation difference alignment loss function is constructed. Specifically, this includes:
[0057] First-order representation difference is defined as the Euclidean distance between the representations of two samples:
[0058]
[0059] in, For encoder, For encoder parameters, and There are two input samples.
[0060] Second-order representational differences are defined as the absolute differences between first-order differences:
[0061]
[0062] To avoid trivial solutions where the differences between all categories collapse to zero, an alignment loss function in the form of relative differences is used:
[0063]
[0064] in, For the sample size, and The weights of the sample pairs calculated in step S1, For sample pairs and The final weights in the difference alignment loss, For sample pairs and The final weights in the difference alignment loss. This formula normalizes the regularization term by the magnitude of the first-order difference to ensure optimization stability while maintaining the discriminative power required for label disambiguation.
[0065] S3: Dimension Masking Stage: To reduce the computational complexity of high-dimensional representation difference alignment and alleviate the problem of distance metric failure in high-dimensional space, a random representation dimension masking strategy is introduced.
[0066] Specifically, let the original representation dimension be... After dimensionality reduction, the dimension is ( Define the mask vector. ,in That is, the mask vector contains One element is 1, and the rest are 0. The difference in the first-order representation after masking is redefined as:
[0067]
[0068] in, This represents element-wise multiplication. This masking strategy preserves discriminative features while effectively reducing computational overhead by randomly emphasizing subsets of the representation dimensions during training.
[0069] S4: Model Integration Stage: Seamlessly integrate the representation difference alignment framework with existing partial label learning methods and iteratively optimize the model parameters.
[0070] Specifically, to ensure compatibility with existing partial labeling learning methods, a unified strategy is adopted: representations of all supervised views (such as pseudo-labels or label distributions) are integrated into a single representation space, where inter-class difference alignment is performed. Therefore, any existing partial labeling learning method can be seamlessly integrated with RDAP without disrupting its original workflow. This method enhances robustness through regularized representation learning while maintaining its advantages in handling complex data.
[0071] The final total loss function is a weighted sum of the original loss from partial label learning and the loss from aligning representation differences:
[0072]
[0073] in, The original loss function for the selected partial label learning method. To balance the hyperparameters of the two losses.
[0074] S5: Classification and prediction stage: Using the neural network model trained in step S4, classify and predict the image examples in the test set to obtain a high accuracy.
[0075] The following experiments demonstrate the advantages of this invention over existing technologies by showcasing representation enhancement of various biased label learning methods on public datasets CIFAR-10, CIFAR-100, and Tiny-ImageNet.
[0076] Two candidate label generation protocols are used: random flipping and instance-related flipping. Seven deep partial label learning methods are evaluated: CC, PRODEN, LWS, CAVL, PaPi, CRDPLL, and PoP. Under this experimental setting, we integrate the proposed RDAP framework with these baseline methods and evaluate classification accuracy on multiple datasets and protocols.
[0077] Experimental results are shown in Tables 1 and 2. Gray rows represent partial label learning methods integrating RDAP, with better results highlighted in bold. The results demonstrate that the proposed method consistently improves the performance of baseline methods in most scenarios. For example, on the CIFAR-100 dataset, for the PRODEN method, the proposed framework achieves performance improvements of 4.11% and 0.93% with label flipping probabilities of 0.1 and 0.2, respectively. Notably, the proposed method demonstrates stronger effectiveness in challenging scenarios, such as high-difficulty datasets or instance-related candidate label generation. For instance, on the Tiny-ImageNet dataset, for the CRDPLL method, the proposed framework achieves improvements of 0.91% and 0.34% with random flipping probabilities of 0.1 and instance-related label generation, respectively.
[0078] Table 1. Classification accuracy (%) on the CIFAR-10 and CIFAR-100 datasets.
[0079]
[0080] Table 2. Classification accuracy (%) on the Tiny-ImageNet dataset
[0081]
[0082] In addition, ablation studies were conducted to verify the effectiveness of the label distribution difference measure based on KL divergence, as well as the threshold. The impact of selection. Experimental results show that using KL divergence as weights is significantly better than the strategy of directly using pseudo-labels. In threshold sensitivity analysis, the method of this invention remains effective under most threshold settings.
[0083] This invention also provides a partial label learning representation enhancement system based on representation difference alignment, comprising: a relation estimation module, which takes as input a partial label dataset consisting of multiple image samples and its candidate label set, and uses a label distribution difference measure based on KL divergence to estimate the visual category relationships between different image samples in a label-independent manner; a difference alignment module, which performs alignment regularization on the visual category representation differences in the high-dimensional image representation space output by the neural network image encoder based on the visual category relationships; a dimensionality masking module, which combines a random representation dimensionality masking strategy to perform dimensionality reduction sampling on the high-dimensional image representation vector to reduce the computational complexity of image representation difference alignment; and a model integration module, which seamlessly integrates the representation difference alignment framework with existing partial label learning image classification methods based on deep neural networks, and iteratively updates the model parameters used for image feature extraction.
[0084] The present invention also provides a storage medium storing a computer program that, when executed by at least one processor, implements the steps of the partial label learning representation enhancement method based on representation difference alignment as described above.
[0085] The present invention also provides an electronic device, including a memory and a processor, wherein: the memory is used to store a computer program that can run on the processor; and the processor is used to execute, when running the computer program, the steps of the partial label learning representation enhancement method based on representation difference alignment as described above.
[0086] It should be understood that the embodiments and descriptions above are only the principles, main features and advantages of the present invention. Various changes and modifications can be made to the present invention without departing from the spirit and scope of the invention, and all such changes and modifications fall within the protection scope of the present invention.
Claims
1. A partial label learning representation enhancement method based on representation difference alignment for image classification, characterized in that, Includes the following steps: (1) Based on a partially labeled dataset consisting of multiple image samples and its candidate label set, the visual class relationship between different image samples is estimated in a label-independent manner using a label distribution difference measure based on KL divergence. (2) Based on the visual category relationships estimated in step (1), the visual category representation differences in the high-dimensional image representation space output by the neural network image encoder are aligned and regularized. (3) Combining the random representation dimension masking strategy, the high-dimensional image representation vector is sampled in a way that reduces the computational complexity of image representation difference alignment; (4) The representation difference alignment framework is seamlessly integrated with the existing partial label learning image classification method based on deep neural networks, and the model parameters used for image feature extraction are iteratively updated to improve the classification accuracy of the image classification model for unknown image samples.
2. The partial label learning representation enhancement method based on representation difference alignment according to claim 1, characterized in that, The implementation process of step (1) is as follows: The label distribution difference measure based on KL divergence is defined as: in, For the sample and The difference in label distribution between them is a metric. Indicates the first Samples during the next training iteration The tags are distributed across categories The value on, Indicates the first Samples during the next training iteration The tags are distributed across categories The value on, For adjustable hyperparameters, Total number of categories; Using confidence threshold Filtering low-confidence samples, the weights of sample pairs in the difference alignment loss are defined as follows: 。 3. The partial label learning representation enhancement method based on representation difference alignment according to claim 1, characterized in that, The visual category representation differences are aligned and regularized using a relative difference form. The regularization loss function is defined as follows: in, For first-order representation differences, For second-order representation differences, For encoder parameters, For the sample size, For sample pairs and The final weights in the difference alignment loss, For sample pairs and The final weights in the difference alignment loss.
4. The partial label learning representation enhancement method based on representation difference alignment according to claim 3, characterized in that, The difference in first-order representations is the Euclidean distance between the representations of the two samples: in, For encoder, For encoder parameters, and There are two input samples; The difference in second-order representations is the absolute difference between first-order differences: 。 5. The partial label learning representation enhancement method based on representation difference alignment according to claim 1, characterized in that, The implementation process of step (3) is as follows: Random representation dimension masking strategy uses mask vectors Implementation, in which , For the original representation dimension, To reduce the dimensionality, the difference in first-order representations after masking is redefined as: in, This indicates element-wise multiplication.
6. The partial label learning representation enhancement method based on representation difference alignment according to claim 1, characterized in that, The implementation process of step (4) is as follows: The representations of all supervised views are integrated into a unified representation space, in which inter-category differences are aligned. The final total loss function is a weighted sum of the original loss from partial label learning and the loss from aligning representation differences: in, The original loss function for the selected partial label learning method. To balance the hyperparameters of the two losses, This is the regularization loss function.
7. A partial label learning representation enhancement system based on representation difference alignment using the method described in any one of claims 1 to 6, characterized in that, include: The relationship estimation module takes a partially labeled dataset consisting of multiple image samples and its candidate label set as input, and uses a label distribution difference measure based on KL divergence to estimate the visual category relationship between different image samples in a label-independent manner. The difference alignment module aligns and regularizes the differences in visual category representations in the high-dimensional image representation space output by the neural network image encoder based on the relationship between visual categories. The dimension masking module, combined with a random representation dimension masking strategy, performs dimensionality reduction sampling on high-dimensional image representation vectors to reduce the computational complexity of image representation difference alignment. The model integration module seamlessly integrates the representation difference alignment framework with existing deep neural network-based partial label learning image classification methods, and iteratively updates the model parameters used for image feature extraction.
8. A storage medium, characterized in that, The storage medium stores a computer program that, when executed by at least one processor, implements the steps of the partial label learning representation enhancement method based on representation difference alignment as described in any one of claims 1 to 6.
9. An electronic device, characterized in that, Includes memory and processor, wherein: Memory is used to store computer programs that can run on a processor; A processor, configured to, while running the computer program, perform the steps of the partial label learning representation enhancement method based on representation difference alignment as described in any one of claims 1 to 6.