Asymmetric dual-task cooperative training biased label learning model and image processing device

Abstract

The application discloses an asymmetric double-task cooperative training partial label learning model and an image processing device, and the model comprises a disambiguation network module, an auxiliary network module and an error correction module, wherein: the disambiguation network module calculates classification probability and label confidence according to sample data; the auxiliary network module calculates final classification probability and label final confidence by using sample data and the classification probability and label confidence generated by the disambiguation network module; and the error correction module is used for carrying out information preliminary extraction and confidence improvement according to the classification results obtained by the disambiguation network module and the auxiliary network module. The effect is that the application proposes a double-task partial label learning model based on deep learning, two networks with the same structure are cooperatively trained through different tasks, the pseudo label recognized by the disambiguation network is used for training, and the error accumulation problem is gradually relieved through information distillation and label confidence fine adjustment.

Description

Technical Field

[0001] This invention belongs to the field of weakly supervised learning technology in artificial intelligence, specifically involving a biased label learning model and image processing device for asymmetric dual-task collaborative training. Background Technology

[0002] In recent years, artificial intelligence (AI) technology has developed rapidly and has been widely applied in various real-world scenarios, such as national defense, healthcare, and e-commerce. Deep learning is a crucial supporting technology for AI, relying on a large number of precisely labeled samples for parameter training. However, in the real world, due to factors such as security, privacy, and sample ambiguity, obtaining accurate standard samples is costly. Therefore, weakly supervised learning methods that rely on incomplete, imprecise, and inaccurate labeled data have received widespread attention in recent years.

[0003] Partially labeled learning is a typical example of weakly supervised learning. Unlike assigning a unique and accurate label to each sample, partially labeled learning assumes that each sample is labeled with a set of candidate labels, of which exactly one is the true label. Assigning a set of candidate labels to a sample is far less difficult than assigning a precise and unique label, especially when the semantic ambiguity of the sample's features is high. For example, Siberian huskies and wolves have similar physical features. For a picture of a Siberian husky or a wolf, accurately identifying its unique true label is difficult for the annotator, resulting in high manpower and time costs. However, assigning a set of candidate labels that includes both Siberian huskies and wolves is much easier.

[0004] Recent research on partially labeled learning has primarily focused on discrimination-based methods, which treat the true labels as latent variables and identify them through label disambiguation. Various algorithms have emerged based on this approach, such as maximum margin methods, graphical model methods, expectation-maximization algorithms, contrastive learning methods, and consistency regularization methods. Among these methods, self-trained deep models are a promising approach, achieving state-of-the-art performance by learning label confidence vectors and iteratively training them.

[0005] However, self-trained partial label learning (PLL) models suffer from error accumulation because complex instances are difficult to classify and are easily misdisambiguated, which can further mislead the model, leading to false positive labels and performance degradation. Cooperative strategies, by training two networks simultaneously and enabling them to interact, offer a viable solution to mitigate error accumulation. While cooperative strategies have been extensively studied in noisy label learning (NLL), they have not been adequately investigated in partial label learning. Recently, Yao et al. (Yao Y, Gong C, Deng J, et al. Network Cooperation with Progressive Disambiguation for Partial Label Learning. In: Proceedings of Machine Learning and Knowledge Discovery in Databases: European Conference, 2020, 471-488) proposed a novel method based on cooperative training, called NCPD. NCPD transforms the partially labeled dataset into a high-noise-rate dataset by data duplication and employs a typical NLL method—co-teaching (Han B, Yao Q, Yu X, et al. Co-teaching: Robust Training of Deep Neural Networks with Extremely Noisy Labels. In: Proceedings of Advances in Neural Information Processing Systems, 2018, 31). However, NCPD not only results in extremely high time and space complexity but also has limited model performance.

[0006] Furthermore, most existing co-training models, including NCPD, are symmetric, meaning their two network branches have the same structure and are trained using the same input data and loss function. They assume that by initializing with different parameters, two structurally identical networks can acquire different capabilities on the same task, thus enabling them to correct each other's errors. However, training in a symmetric mode makes the two networks more prone to the same problems; for example, both networks may struggle to correctly identify complex samples. Therefore, they cannot effectively correct errors. Summary of the Invention

[0007] In view of this, the primary objective of this invention is to provide an asymmetric dual-task collaborative training partial label learning model. By collaboratively training two networks with the same structure on different tasks, the model parameters can be effectively trained from partially labeled data to obtain a classifier for the test environment of partially labeled datasets, and it exhibits superior performance under uniform partial label generation strategies and instance-dependent partial label generation strategies.

[0008] To achieve the above objectives, the specific technical solution adopted by the present invention is as follows:

[0009] A partial label learning model for asymmetric dual-task collaborative training is characterized by comprising a disambiguation network module, an auxiliary network module, and an error correction module, wherein:

[0010] The disambiguation network module is used to calculate the classification probability and label confidence of each category based on the sample data;

[0011] The auxiliary network module takes the sample data and the classification probability and label confidence generated by the disambiguation network module as input, and outputs the final classification probability and label confidence of the sample data in each category.

[0012] The error correction module is used to perform preliminary information extraction and confidence improvement based on the classification results obtained by the disambiguation network and the final classification results obtained by the auxiliary network module. The preliminary information extraction is used to treat the final classification probability obtained by the auxiliary network module as the true distribution and introduce a loss based on KL divergence to ensure that the predicted probability of the disambiguation network module is consistent with the predicted probability of the auxiliary network module. The confidence improvement is used to fine-tune the label confidence of the disambiguation network module based on the final label confidence obtained by the auxiliary network module.

[0013] Optionally, the disambiguation network module calculates the classification score using a multilayer perceptron and then calculates the classification probability p using a Softmax function. i ∈R m , where p i The k-th element p in ik Represents sample data x i The probability of being classified into label k;

[0014]

[0015] Where m is the total number of tag categories, y i For sample data x i Class results, MLP k (x i ) indicates that the data x i The score for class k is assigned to the label, and the temperature parameter is τ.

[0016] Optionally, the sample data is image data, and two different image enhancement techniques are used for each sample image x. i Create two enhanced images, denoted as x′ i =Aug1(x i ) and x″ i =Aug2(x i ), and form an augmented sample dataset (x i )={x i ,x′ i ,x″ i};

[0017] The disambiguation network module is based on the enhanced sample dataset. The classification probability p′ corresponding to the augmented sample is calculated. i ,p″ i and label confidence c′ i ,c″ i .

[0018] Optionally, the disambiguation network module:

[0019] according to Calculate sample data x i The classification consistency loss, where, express The model, x represents i The sample set consisting of its augmented samples, express Any sample in, Indicates sample The probability of classifying a class as category k, Y i Represents sample data x i The candidate tag set;

[0020] according to Calculate sample data x i The overall confidence score of class k is used to obtain the sample data x. i Category-based comprehensive confidence vector w i =[w i1 ,w i2 ,…,w im ],in Indicates sample x i Corresponding sample set any sample in The confidence score on category k, which is based on the sample Probability prediction results from the previous round Calculate if category k∈Y i ,but otherwise

[0021] according to Calculate sample data x i The risk of consistent loss, among which express The model, x represents i The sample set consisting of its augmented samples, w ik Indicates sample x i The confidence score for belonging to category k. express Any sample in, Indicates sample The probability of classifying a class as category k, Y i Represents sample data x i The candidate label set is obtained. By minimizing the risk consistency loss after data augmentation, the classification probabilities of both the original sample data and the augmented sample data gradually approach the comprehensive confidence vector w. i ;

[0022] Optionally, the disambiguation network module is configured according to the constructed loss function L. disam (x i ) = L cc (x i )+γ(t)L rc (x i Training is performed, where the weight parameters λ and T are hyperparameters, and t is the number of training epochs.

[0023] Optionally, pseudo-class labels are generated for each instance based on the prediction results of the disambiguation network. Pseudo-class labels From the confidence vector c i The label with the highest probability is selected, representing the most likely true label according to the model. If a pair of sample data shares the same pseudo-label, they are assigned a similarity label of 1; otherwise, a similarity label of 0 is assigned. The generated similarity dataset is represented as follows. in, They are used to represent sample x respectively i and x j Feature representation extracted by the auxiliary network Let χ represent a pair of samples, where χ represents the sample space, and s represents the sample space. ij ∈{0,1} represents the result of sample x i and x j The similarity labels of the formed sample pairs are used to train the auxiliary network module using the generated similarity dataset;

[0024] For sample data pairs with a similarity label of 1, it is expected that their predicted classification probabilities will exhibit high similarity. The loss function used during training is:

[0025] Among them, the binary cross-entropy loss function Indicates sample x i The classification probability of class k is predicted by the auxiliary network. Indicates sample x i The classification probability of a data-enhanced sample in class k. This indicates the percentage of time the stopping loss function is used during backpropagation. Gradient updates.

[0026] Optionally, the KL divergence-based loss introduced in the initial information extraction step is calculated according to:

[0027] Calculate; where KL(·) represents the KL divergence function.

[0028] Optionally, in the confidence improvement step, the fine-tuned confidence is calculated in the t-th round as follows:

[0029] in, Sample x for auxiliary network calculation i The overall confidence vector, μ(t) = min(ρ×max(tt) 0, 0),μ max ) is an increasing function of training epochs t, where t0 and t0 are the training epochs

[0030] μ max It is a hyperparameter; before the t0th training iteration, μ(t) = 0, and 0 ≤ μ max ≤1 is the upper limit of μ(t), the growth rate of μ depends on ρ, and during co-training, the original confidence level w i Improved confidence level Replacement.

[0031] Optionally, according to The model is trained by setting the overall training loss. First, the disambiguation network module is preheated to ensure that the true labels of some training sample data are accurately identified through the confidence vector. Then, the pre-trained model parameters are used to initialize the parameters of the auxiliary network. In addition, in order to enhance the efficiency of model training, noisy similarity labels are generated using sample data within the same mini-batch. In the inference phase, one of the modules, either the disambiguation network module or the auxiliary network module, is selected for prediction.

[0032] Based on the above model, the present invention also provides an image processing device, the key of which is: to perform image classification using the asymmetric dual-task collaborative training biased label learning model described above.

[0033] The beneficial effects of this invention are:

[0034] (1) This invention explores asymmetric collaborative training of partial label learning and proposes a novel deep learning-based dual-task partial label learning model, which trains two networks with the same structure through different tasks.

[0035] (2) This invention proposes an effective supervised learning auxiliary network, which is trained using pseudo-labels identified by a disambiguation network, and gradually alleviates the problem of error accumulation through information distillation and label confidence fine-tuning.

[0036] (3) Extensive experimental results on benchmark datasets show that the proposed model exhibits superior performance under both uniform partial label generation strategies and sample data-dependent partial label generation strategies. Attached Figure Description

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0038] Figure 1 This is a diagram of a partial label learning model architecture for asymmetric dual-task collaborative training provided by the present invention;

[0039] Figure 2 This is a schematic diagram of generating noise pairwise similarity labels based on the confidence vector in a specific embodiment of the present invention. Detailed Implementation

[0040] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the preferred embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0041] like Figure 1 As shown, this invention provides an asymmetric dual-task collaborative training partial label learning model, including a disambiguation network module, an auxiliary network module, and an error correction module, wherein:

[0042] The disambiguation network module is used to calculate the classification probability and label confidence of each category based on the sample data;

[0043] The auxiliary network module takes the sample data and the classification probability and label confidence generated by the disambiguation network module as input, and outputs the final classification probability and label confidence of the sample data in each category.

[0044] The error correction module is used to perform preliminary information extraction and confidence improvement based on the classification results obtained by the disambiguation network and the final classification results obtained by the auxiliary network module. The preliminary information extraction is used to treat the final classification probability obtained by the auxiliary network module as the true distribution and introduce a loss based on KL divergence to ensure that the predicted probability of the disambiguation network module is consistent with the predicted probability of the auxiliary network module. The confidence improvement is used to fine-tune the label confidence of the disambiguation network module based on the final label confidence obtained by the auxiliary network module.

[0045] In practice, the disambiguation network module and the auxiliary network module have the same network architecture, and the model training process mainly includes the following steps:

[0046] Step 1: Data Preparation and Preprocessing

[0047] First, extract d-dimensional features from the dataset. Assume the dataset is D = {(x...} i ,Y i )}, where x i Y is a feature of the i-th sample. i It is its candidate label set, which labels each training sample as a candidate label set, that is, for each sample x i , corresponding to a label set Y i ,in C represents the total number of categories.

[0048] Data augmentation is performed on the original data to generate multiple views, forming an instance collection. Specifically, for each sample x... i Multiple variant instances are generated using data augmentation techniques (such as rotation, cropping, and color dithering). Data augmentation helps improve the model's generalization ability and robustness. Taking image data as an example, in practice, two different image augmentation techniques can be used for each sample image x. i Generate two enhanced images, denoted as x′. i =Aug(x) i ) and x″ i =Aug(x) i ), and constitute an augmented sample dataset

[0049] Step 2: Train the disambiguation network

[0050] First, the disambiguation network module calculates the classification score using a multilayer perceptron, and then calculates the classification probability p using a softmax function. i ∈R m , where p i The k-th element p in ik Represents sample data x i The probability of being classified into label k is determined by:

[0051] calculate,

[0052] Where m is the total number of tag categories, y i For sample data x i The classification results, MLP k (x i ) indicates that the sample data x i The classification score is assigned to label k, and τ is the temperature parameter.

[0053] Next, the label confidence vector is calculated. The disambiguation network module calculates the label confidence vector based on the augmented sample dataset. The classification probability p′ corresponding to the augmented sample is calculated. i ,p″ i and label confidence c′ i ,c″ i .

[0054] Then follow Calculate instance data x i The classification consistency loss, where, express The model, x represents i The sample set consisting of its enhanced instances, express Any sample in, Representation of instances The probability of classifying a class as category k, Y i Represents instance data x i The candidate tag set.

[0055] In order Calculate instance x i The combined confidence score of each category k is used to obtain instance x. i Category-based comprehensive confidence vector w i =[w i1 ,w i2 ,…,w im ].in Indicates sample x i Corresponding sample set any sample in The confidence score on category k, which depends on the instance. Probability prediction results from the previous round Calculate if category k∈Y i ,but otherwise

[0056] Finally according to Calculate instance xi The risk of consistent loss, among which, express The model, x represents i The sample set consisting of its augmented samples, w ik Indicates sample x i The confidence score for belonging to category k. express Any sample in, Representation of instances The probability of classifying a class as category k, Y i Represents sample data x i The candidate label set is calculated. By minimizing the risk consistency loss after data augmentation, the classification probabilities of both the original instance data and the augmented instance data gradually approach the comprehensive confidence vector w. i .

[0057] Step 3: Generate low-noise similarity labels

[0058] First, pseudo-class labels are generated. Based on the confidence levels learned by the disambiguation network, pseudo-class labels are generated for each instance. From the confidence vector c i The label with the highest probability is selected, representing the label that the model considers most likely to be the true label.

[0059] Next, the pseudo-class labels are converted into pairwise similarity labels, resulting in low-noise pairwise similarity labels. Specifically, if a pair of sample data shares the same pseudo-label, then... If a sample is a positive pair, it is assigned a similarity label of 1, indicating that the two samples belong to the same class; otherwise, it is a negative pair, assigned a similarity label of 0, indicating that the two samples belong to different classes. The generated similarity dataset is represented as follows: in, They are used to represent instance x respectively i and x j Feature representation extracted by the auxiliary network Let χ represent a pair of instances, where χ represents the instance space, and s represents the instance space. ij ∈{0,1} represents the result of sample x i and x j The similarity labels of the formed sample pairs are used to train the auxiliary network module using the generated similarity dataset;

[0060] Step 4: Train the model network

[0061] First, calculate the loss function for the auxiliary network. The loss function used during the training of the auxiliary model is: Wherein, the binary cross-entropy loss function is Representing instance x i The classification probability of class k is predicted by the auxiliary network. Representing instance x i The classification probability of the data augmented instance in category k. This indicates the percentage of time the stopping loss function is used during backpropagation. Gradient updates.

[0062] Next, the loss based on KL divergence is calculated. According to... Calculate the KL divergence loss, where KL(·) represents the KL divergence function.

[0063] Then, the confidence level is refined. In round t, the fine-tuned confidence level is calculated as follows:

[0064] in, Sample x for auxiliary network calculation i The overall confidence vector, μ(t) = min(ρ×max(tt) 0, 0),μ max ) is an increasing function of training epochs t, where t0 and μ are... max It is a hyperparameter; before the t0th training iteration, μ(t) = 0, and 0 ≤ μ max ≤1 is the upper bound of μ(t), and the growth rate of μ depends on ρ. During co-training, the original confidence vector w i Improved confidence level Replacement.

[0065] Finally according to Set the overall training loss of the model and train it.

[0066] Step 5: Model Inference

[0067] During the inference phase, a disambiguation network is used for prediction. That is, the input is a test sample x, and the output is a classification result. The confidence scores provided by the auxiliary network are used to refine the results, further improving prediction accuracy. The outputs of both networks are combined to obtain the final prediction result, thereby improving the overall performance of the model.

[0068] In the specific implementation process, an image processing device is also provided, which can perform image classification using the asymmetric dual-task collaborative training biased label learning model described above.

[0069] To verify the effectiveness of this invention, experiments were conducted on multiple benchmark datasets, including SVHN, CIFAR-10, and CIFAR-100. The classification accuracy of this invention was compared with that of several state-of-the-art partial label learning methods, demonstrating its superiority. The comparison methods mainly include: RC (Feng L, Lv J, Han B, et al. Provably Consistent Partial-Label Learning. In: Proceedings of Advances in NeuralInformation Processing Systems, 2020, 33: 10 948-10 960.), CC ( Feng L, Lv J, Han B, et al. Provably Consistent Partial-Label Learning. In: Proceedings of Advances in Neural Information Processing Systems, 2020, 33: 10948-10 960.), PRODEN (Lv J, Xu M, Feng L, et al. Progressive Identification of True Labels for Partial-Label Learning. In: Proceedings of the 37th International Conference on Machine Learning, 2020, 6500-6510.), PiCO (Wang H, Xiao R,Li Y,et al.Contrastive LabelDisambiguation for Partial Label Learning.In:Proceedings of the International Conference on Learning Representations,2022.), DPLL(Dong-Dong Wu,Deng-BaoWang,and Min-Ling Zhang.2022.Revisiting consistency regularization for deeppartial label learning.In Proceedings of the International Conference onMachine Learning.PMLR, 24212-24225), NCPD (Yao Y, Gong C, Deng J, et al. Network Cooperation with Progressive Disambiguation for Partial Label Learning. In: Proceedings of Machine Learning and Knowledge Discovery in Databases: European Conference, 2020, 471-488.), etc. .

[0070] This implementation is based on PyTorch and uses an 18-layer ResNet as the backbone network to encode features on image datasets (including SVHN, CIFAR-10, and CIFAR-100). Since CNAE-9 and BirdSong are not very large, this example constructs linear layers as feature encoders for these two non-image datasets. For these two non-image datasets, we increase the amount of data in the example by adding random labels and Gaussian noise. The model is optimized using the SGD optimizer with a momentum value of 0.9 and a weight decay of 1e-4. The initial learning rate is set to 0.1, and it is reduced to one-tenth of its original value at epochs 100 and 150. The total number of training epochs is set to 200, with 50 warm-up epochs on CIFAR-100 when q = 0.2, and 20 in other cases. The batch size is set to 64. The τ value is searched within the range [1, 5, 10, 20, 30], and finally τ = 20 is chosen. When calculating λ, the T value and λ... max The values ​​were set to 100 and 1, respectively. Furthermore, the confidence refinement-related hyperparameter ρ was set to 0.02, t0 was determined by adding the number of warm-up epochs to 50, and μ... max Set it to 0.9.

[0071] The following are specific examples and their experimental results:

[0072] The data for this example comes from three benchmark datasets: SVHN (Netzer Y, Wang T, Coates A, et al. Reading Digits in Natural Images with Unsupervised Feature Learning. InNIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2011.), CIFAR-10, and CIFAR-100 (Krizhevsky A. Learning multiple layers of features from tiny images. Computer Science, 2009.).

[0073] The data construction in this example first constructs partially labeled datasets for these four benchmark datasets through two generation processes: uniform generation and instance-dependent generation. The method for generating the datasets comes from (Wu DD, Wang DB, Zhang ML. Revisiting Consistency Regularization for Deep Partial Label Learning. In: Proceedings of the 39th International Conference on Machine Learning, 2022, 162:24 212-24225.). In the uniform generation process, the probability of an incorrect label becoming a candidate label is the same, denoted as q. q takes values ​​{0.1, 0.3, 0.5, 0.7} on SVHN, CIFAR-10, and CNAE-9, and {0.01, 0.05, 0.1, 0.2} on CIFAR-100. In the instance-dependent candidate label set generation process, an 18-layer ResNet (ResNet-18) is pre-trained. The probability of an incorrect label j becoming a false positive label is calculated as follows:

[0074]

[0075] Where g′ j (x i ) is the given input x i Then, the classification probability of label j is calculated using the pre-trained ResNet-18.

[0076] This example uses the same backbone network, learning rate, optimizer, and batch size across all comparison methods (including supervised learning). For methods that did not initially utilize data augmentation techniques (e.g., RC, CC, and PRODEN), we augment these models by introducing data augmentation methods. To obtain reliable and robust results, we conducted three replicate experiments with different random seeds and report the mean and standard deviation of these results. All models were trained on a GeForce RTX 3090 with 24GB of memory, except for NCPD when q=0.7 on the SVHN dataset. Since NCPD exhibits a memory limit error (MLE) on the SVHN dataset when q=0.7, we ran the experiments on a V100 with 32GB of memory to obtain the results.

[0077] On three different datasets, this embodiment demonstrates a significant performance advantage for data with different q values. For example, on CIFAR-10, when q is 0.1, 0.3, 0.5, and 0.7, the proposed model outperforms the optimal model by 0.361%, 0.625%, 0.638%, and 1.694%, respectively. Except for a slightly lower performance than NCPD on the SCHN dataset with q=0.1, the proposed model significantly outperforms NCPD on all other datasets. This result further demonstrates the effectiveness of asymmetric dual-task collaborative training.

[0078] The method of this invention is highly competitive with supervised learning. For example, on CIFAR-10 with q=0.1, and on CIFAR-100 with q=0.01 and q=0.05, the method of this invention even outperforms supervised learning. Further, removing data augmentation from both the method of this invention and supervised learning, the results show that the method of this invention still significantly outperforms supervised learning. This indicates that the method of this invention can effectively combine the different performances of the two networks through co-training, mining high-quality supervisory information from partially labeled data.

[0079] The method of this invention exhibits strong robustness in terms of data quality. As the q-value increases, the performance of most methods significantly decreases. For example, on CIFAR-10, when q increases from 0.1 to 0.7, the accuracy of comparative methods decreases by 2% to 20%, while the accuracy of the method of this invention fluctuates only within a narrow range of 96.645% to 95.550%. Therefore, it can be concluded that the accuracy gap between the method of this invention and the state-of-the-art methods becomes more pronounced as the q-value increases. On SVHN, when q is 0.5 and 0.7, the performance improvement of the method of this invention is 0.228% and 1.162%, respectively. Similarly, on CIFAR-100, when q is 0.01 and 0.2, the accuracy improvement of the method of this invention is 1.475% and 1.745%, respectively.

[0080] In this embodiment, the method of the present invention not only performs excellently on uniformly generated partially labeled datasets, but also outperforms other models on instance-dependent partially labeled datasets. For example, it shows a significant performance improvement of 1.290% on CIFAR-10 and an accuracy improvement of 0.895% on SVHN. This advantage further validates the capability and generalization ability of the method of the present invention.

[0081] Through the above embodiments, it can be seen that the method model of the present invention exhibits excellent classification performance on different types of datasets, significantly improving classification accuracy, thus proving the effectiveness and wide applicability of the present invention in solving the problem of partial label learning.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A partial label learning model for asymmetric dual-task collaborative training, characterized in that, It includes a disambiguation network module, an auxiliary network module, and an error correction module, wherein: The disambiguation network module is used to calculate the classification probability and label confidence of each category based on the sample data, which is image data; The auxiliary network module takes the sample data and the classification probability and label confidence generated by the disambiguation network module as input, and outputs the final classification probability and label confidence of the sample data in each category. The error correction module is used to perform preliminary information extraction and confidence improvement based on the classification results obtained by the disambiguation network and the final classification results obtained by the auxiliary network module. The preliminary information extraction is used to treat the final classification probability obtained by the auxiliary network module as the true distribution and introduce a loss based on KL divergence to ensure that the predicted probability of the disambiguation network module is consistent with the predicted probability of the auxiliary network module. The confidence improvement is used to fine-tune the label confidence of the disambiguation network module based on the final label confidence obtained by the auxiliary network module.

2. The asymmetric dual-task collaborative training partial label learning model according to claim 1, characterized in that, The disambiguation network module calculates the classification score using a multilayer perceptron and then calculates the classification probability using a softmax function. ,in The first in element Representing sample data Categorized into tags Probability, according to: calculate, Where m is the total number of tag categories. For sample data The classification results Indicates sample data Categorize to Tags Category score, This refers to the temperature parameter.

3. The asymmetric dual-task collaborative training partial label learning model according to claim 1 or 2, characterized in that, Two different image enhancement techniques were used for each sample image. Generate two enhanced images, denoted as... and And constitute an augmented sample dataset ; The disambiguation network module is based on the enhanced sample dataset. Calculate the classification probability corresponding to the augmented sample. and label confidence .

4. The asymmetric dual-task collaborative training partial label learning model according to claim 3, characterized in that, The disambiguation network module: according to Calculate samples The classification consistency loss, where, express The model, express The sample set consisting of its augmented samples, Indicates sample Categorize into Category The probability, Representing sample data The candidate tag set; In order Calculate sample data Belongs to various categories The overall confidence level is used to obtain the sample data. Category-based comprehensive confidence vector ;in Indicates sample Corresponding sample set any sample in In category The confidence level is based on the sample. Probability prediction results from the previous round Calculate, if category , but ,otherwise ;according to Calculate samples The risk of consistent loss, among which, express The model, express The sample set consisting of its augmented samples, Indicates sample Category Confidence level, Indicates sample Categorize into Category The probability, Representing sample data The candidate label set is obtained; by minimizing the risk consistency loss after data augmentation, the classification probabilities of both the original sample data and the augmented sample data gradually approach the comprehensive confidence vector. .

5. The asymmetric dual-task collaborative training partial label learning model according to claim 4, characterized in that: The disambiguation network module follows the constructed loss function. Among them, weight parameters , It's a hyperparameter. .

6. The asymmetric dual-task collaborative training partial label learning model according to claim 5, characterized in that: Based on the prediction results of the disambiguation network, pseudo-class labels are generated for each instance; pseudo-class labels From the confidence vector The label with the highest probability is selected, representing the label that the model considers most likely to be the true label. If a pair of sample data shares the same pseudo-label, then assign them a similarity label of 1; otherwise, assign a similarity label of 0, and the generated similarity dataset is represented as follows. ,in, Used to represent samples and Feature representation extracted by the auxiliary network This represents a pair of samples. Representing the sample space, Indicates that the sample and The similarity labels of the formed sample pairs are used to train the auxiliary network module using the generated similarity dataset; For sample data pairs with a similarity label of 1, it is expected that their predicted classification probabilities will exhibit high similarity. The loss function used during training is: ; Wherein, the binary cross-entropy loss function is ; Indicates sample The category predicted by the auxiliary network The classification probability on Indicates sample Data augmentation samples in categories The classification probability on; This indicates the percentage of time the stopping loss function is used during backpropagation. Gradient updates.

7. The asymmetric dual-task collaborative training partial label learning model according to claim 6, characterized in that: The loss based on KL divergence introduced in the initial information extraction step is as follows: Calculate; where, This represents the KL divergence function.

8. The asymmetric dual-task collaborative training partial label learning model according to claim 7, characterized in that: In the confidence improvement step, at the first... The confidence level after fine-tuning is calculated for each round as follows: ,in, Samples for auxiliary network computation The overall confidence vector, Training rounds an increasing function, and It's a hyperparameter, in the first... Before each training session ,and yes The upper limit, The growth rate depends on During collaborative training, the original confidence vector Improved confidence level Replacement.

9. The asymmetric dual-task collaborative training partial label learning model according to claim 8, characterized in that: according to The model is trained by setting the overall training loss. First, the disambiguation network module is preheated to ensure that the true labels of some training sample data are accurately identified through the confidence vector. Subsequently, the pre-trained model parameters are used to initialize the parameters of the auxiliary network. In addition, to enhance the efficiency of model training, noisy similarity labels are generated using sample data within the same mini-batch. During the inference phase, either the disambiguation network module or the auxiliary network module is selected for prediction.

10. An image processing apparatus, characterized in that: Image classification is performed using the asymmetric dual-task collaborative training partial label learning model as described in any one of claims 1-9.

Images