An anti-noise label image recognition method based on meta learning

Abstract

The application discloses an anti-noise label image recognition method based on meta learning, which comprises the following steps: (1) dataset construction: constructing a verification dataset, a training dataset and a to-be-recognized image dataset; (2) noise label correction: innovatively fusing an initial data label, a candidate correction label and an image recognition model prediction label, generating an aggregated correction label by comprehensively integrating the information of the three through a weight coefficient to replace the original noise label; then, a gradient descent algorithm is used to update the noise label correction model to reduce the loss of the verification dataset to the maximum extent; and (3) meta learning parameter optimization: constructing a joint optimization target containing image recognition model parameters and noise label correction model parameters, and realizing the collaborative gradient update of the two parameter spaces through a double optimization algorithm. The application can be widely applied to image recognition datasets, especially the recognition of image datasets containing noise labels disclosed on the network.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a noise-resistant label image recognition method based on meta-learning. Background Technology

[0002] In today's digital age, deep neural networks, as one of the key technologies in the field of artificial intelligence, have demonstrated significant advantages in image recognition. For example, facial recognition technology, through deep analysis and learning of large amounts of facial image data using deep neural networks, can highly accurately identify different individuals and is widely used in fields such as security monitoring and access control.

[0003] However, in practical applications, deep neural networks face the problem of training data bias. Much training data comes from sources such as web searches and crowdsourcing, and the quality of these annotations varies greatly. For example, images obtained from the internet may have incorrect original labels for various reasons. This prevalence of biased training data can negatively impact the performance and accuracy of deep neural networks.

[0004] Currently, the main methods for addressing this problem fall into two categories: sample selection and label correction. Sample selection methods address this challenge by assigning importance weights to samples one by one based on their loss values. Some works have made the weighting schemes more adaptive by guiding the network training process with a small amount of validation data. These methods aim to discard suspicious noisy samples during training, but even if most of the discarded samples are noisy, they often contain useful information that can improve the accuracy and robustness of the network, especially under high noise ratio conditions.

[0005] Label correction methods mitigate this problem by attempting to find and correct noisy labels to their potential true labels. Some works attempt to estimate the noise transition matrix, i.e., the probability mapping from true labels to noisy labels, and then use the estimated matrix to correct corrupted samples. However, the matrix size grows exponentially with the number of classes, making noisy label correction complex and tractable on large-scale datasets. Furthermore, these methods assume that the noise is class-related, and perform poorly for more complex noise, such as feature-related noise.

[0006] Therefore, researching how to robustly train image recognition models in training data with noisy labels is not only of significant academic research value, but also of great importance to promoting the practical application of deep learning in more fields. Summary of the Invention

[0007] To address the shortcomings of existing technologies, such as Figure 1As shown, this invention designs a noise-resistant label image recognition method based on meta-learning from the perspective of meta-learning, which automatically corrects noise labels step by step in an iterative manner.

[0008] The objective of this invention is achieved through the following technical solution: a method for noise-resistant label image recognition based on meta-learning, comprising the following steps:

[0009] S1. Construct the validation dataset, training dataset, and image dataset to be recognized;

[0010] S2. Input a small batch of training data from the training dataset into the image recognition model and output the predicted labels;

[0011] S3. Input the predicted labels and the initial labels from the training data into the noise label correction model to generate candidate correction labels;

[0012] S4. Employ a dynamic weighting strategy to weight and fuse the predicted label, the initial label of the training data, and the candidate correction label to generate an aggregated correction label.

[0013] S5. Construct a meta-learning parameter optimization strategy, which includes inner-layer optimization and outer-layer optimization.

[0014] S6. Select the next batch of training data from the training dataset and repeat steps S2 to S5 until the parameters of the two models converge. Use the image recognition model with converged parameters as the application image recognition model, input the image dataset to be recognized into the application image recognition model, and output the final image recognition result.

[0015] Further, step S1 specifically includes: selecting a small amount of clean data from the noisy dataset to form a validation dataset, and using the remaining data as a training dataset, as expressed by the following mathematical formula:

[0016]

[0017] in Let x represent the training dataset, n represent the number of samples in the training dataset, and x represent the training dataset. i Let y represent the i-th training data in the training dataset. i This represents the label corresponding to the i-th training data. Let m represent the validation dataset, where m represents the number of samples in the validation dataset (which is much smaller than the number of samples in the training dataset). This represents the i-th validation data in the validation dataset. This represents the label corresponding to the i-th validation data. The validation dataset contains only clean labels and is used for training the noise label correction model; the training dataset contains both clean and noisy labels and is used for training the image recognition model.

[0018] Further, step S4 specifically includes: inputting the noise label, the predicted label, and the candidate corrected label into the noise label correction model, and outputting the aggregated corrected label, as expressed by the mathematical formula below:

[0019]

[0020] in Let represent the aggregated corrected labels, t represent the training period, g represent the noise label correction model, and y represent the initial labels of the training data (which may be noise labels). This represents the predicted label output by the image recognition model. This represents the candidate correction label (an aggregated correction label from the previous training cycle, used to mitigate the negative impact of significant fluctuations in the prediction results of the image recognition model during training, and to improve the accuracy of the correction label), w (t) These represent the parameters of the noise label correction model.

[0021] Specifically, y, The final aggregated correction label is obtained by combining the weights, and the mathematical formula is as follows:

[0022]

[0023] in This represents the aggregated correction label, and α represents the weight value.

[0024] Further, step S5 specifically includes: using the aggregated corrected labels generated by the noise label correction model and the validation dataset, jointly training the image recognition model and the noise label correction model through a meta-learning parameter optimization strategy, as expressed by the following mathematical formula:

[0025]

[0026] Where θ * θ represents the optimal parameters of the image recognition model. * (w) indicates that there is a nested relationship between parameters w and θ, where θ represents the parameters of the image recognition model and w represents the parameters of the noise correction model. w represents the loss of the training dataset. * The optimal parameters of the noise label correction model are represented. Let represent the validation dataset loss, f represent the noise label correction model, and l represent the loss function, such as the cross-entropy loss function. When the noise label correction model parameters reach their optimum, i.e., w = w... * It tends to correct noisy labels to clean labels, thereby greatly improving the training quality of subsequent image recognition models.

[0027] The parameters w and θ in the above formulas (5) and (6) have a nested relationship. * (w), meaning that for each step of gradient descent, parameter θ must also be updated synchronously. Furthermore, the computational relationship between the two parameters at each step needs to be stored until both are updated to their optimal values. Such nested parameter optimization involves significant computational resource consumption and is almost impossible to implement in practice. Therefore, this method utilizes local gradient information, that is, using the parameters after a single step of gradient descent to approximate the optimal parameters for updating. The specific steps are as follows:

[0028] First, let the mini-batch training samples in the training dataset be denoted as Training dataset loss It can be represented as: Furthermore, the formula for updating parameter θ with respect to parameter w in single-step gradient descent can be derived:

[0029]

[0030] in This represents the intermediate result of updating parameter θ. (t) The parameters of the image recognition model at training period t are represented by η1, which represents the learning rate. Let represent the derivative function. Then, a mini-batch of validation samples from the validation dataset is used. Update the parameter w using single-step gradient descent:

[0031]

[0032] Among them, w (t+1) Let represent the updated noise correction model parameters, and η² represent the learning rate. Finally, using w... (t+1) Update parameter θ:

[0033]

[0034] Where, θ (t+1) This represents the updated image recognition model parameters.

[0035] Compared with the prior art, the present invention has the following improvements and advantages:

[0036] 1. This invention abandons the traditional predefined generation rules and can automatically map noise labels to corrected labels. This approach greatly improves the flexibility of the noise label correction process, enabling it to better adapt to complex real-world image recognition datasets with diverse types and varying noise levels.

[0037] 2. During the iteration process, the present invention uses historical correction information and adaptive weights to gradually optimize the noise label corrector, thereby effectively reducing the accumulation of erroneous information caused by noise labels.

[0038] 3. This invention does not depend on a specific model structure, has strong versatility, and can be easily applied to mainstream image recognition models such as CNN and Transformer. Attached Figure Description

[0039] Figure 1 This is a diagram illustrating the method architecture of the present invention;

[0040] Figure 2 A flowchart of the steps provided for embodiments of the present invention;

[0041] Figure 3 This is a structural diagram of the noise label correction model provided in an embodiment of the present invention;

[0042] Figure 4 A comparison chart of recognition accuracy calculation results for noisy CIFAR-10 and CIFAR-100 datasets provided in this embodiment of the invention;

[0043] Figure 5 A heatmap showing the calculation of noise label correction accuracy on a CIFAR-10 dataset containing 40% symmetrical noise, provided for embodiments of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0045] This application provides a noise label correction method based on meta-learning, and the specific implementation process is as follows: Figure 2 As shown, this covers steps S1 to S9. It should be noted that... Figure 2 The illustrated execution order is merely one possible sequence presented in the embodiments of this application, and is not the only possible execution order of the method. Where the final result can be achieved, the steps can be executed in parallel or in reverse order. Furthermore, Figure 3 The noise label correction model structure shown and the image recognition model structure described in this paper are general examples presented in the embodiments of this application, and are not the only structures. In practical applications, they can be flexibly adjusted and optimized according to specific needs and scenarios.

[0046] S1: Create the dataset. The dataset includes a training dataset, a test dataset, and a validation dataset. The training dataset includes noisy labels, the validation dataset includes only clean labels, and the test dataset is the dataset of images to be recognized.

[0047] Specifically, the CIFAR-10 dataset can be downloaded from the official website. It contains 60,000 images across 10 classes. 50,000 images are used for training, 10,000 for testing, and 100 images are randomly selected from each class in the training dataset as the validation dataset. This ensures the validation dataset is representative of all classes in the training dataset, but with a much smaller sample size. The training dataset is represented as follows: It consists of n samples, for each image x i Its corresponding label is y i Two noise types were selected to add noise to the labels of the training dataset, with the noise ratio set to p: (1) Symmetric noise: A clean label is added with probability p. Convert to other tags, and with (1) The probability of keeping the original label is p, where k is the number of labels. (2) Asymmetric noise: A clean sample label is corrupted to another label with probability p, and the original label is kept with probability 1-p. The validation dataset is represented as follows: It consists of m clean image samples, which is much smaller than the number of training datasets. Its main purpose is to help the noise label correction model learn the correct semantics.

[0048] S2: Input noisy data into the image recognition model and output predicted labels.

[0049] Before inputting data into the image recognition model, the image recognition model is pre-trained and fine-tuned on the training dataset to enable it to represent noisy image features.

[0050] Specifically, the image recognition model consists of ResNet (a deep convolutional neural network) and is pre-trained on ImageNet (an image dataset), as follows: Where x represents the data input into the model, and θ is the parameter of the image recognition model. This is the output prediction result. After fine-tuning and pre-training, the training data is input into the image recognition model, according to the formula... Calculate the predicted label for the sample.

[0051] S3: Input the initial labels corresponding to the training data and related auxiliary inputs into the noise label correction model (structure as follows) Figure 3 As shown), the aggregated correction label is obtained.

[0052] Specifically, the predicted labels obtained in step S2 Initial label y i and candidate correction labels The input is fed into the noise label correction model, and the aggregated corrected label is obtained after combining the weights. It's important to note here that when y=1, i.e., in the first training cycle, there's a lack of candidate correction labels. At this point, the settings... The predicted labels of the image recognition model are used instead of the candidate correction labels in the calculation.

[0053] S4: Calculate the loss using the predicted labels and the aggregated corrected labels, and calculate the gradient of the loss with respect to the image recognition model parameters. Specifically, after obtaining the predicted labels and the aggregated corrected labels, calculate the gradient of the loss with respect to the parameters of the image recognition model according to the formula... Calculate the training set loss. Specifically, the cross-entropy loss function is chosen. The cross-entropy loss function converges to the optimal solution faster during backpropagation, effectively avoiding the gradient vanishing problem, thus ensuring the stability and efficiency of model training.

[0054] Furthermore, due to training set loss It covers two different gradients: noise label correction model and image recognition model, according to the formula Calculate the gradient of the training set loss with respect to the image recognition model parameters θ. This gradient plays a crucial role in subsequent updates to the image recognition model parameters, reflecting the direction of the model's error on the current training set and enabling the model to optimize in the direction of reducing error.

[0055] S5: Update the image recognition model parameters while maintaining the gradient computation graph.

[0056] Specifically, due to the nested relationship between parameters w and θ, θ * Therefore, when updating the parameter θ, it is necessary to completely save the calculation relationship between the two at each step to ensure the coherence and accuracy of the entire model's calculation logic. To reduce the aforementioned storage and computational resources, a one-step gradient descent approach is used to update the image recognition model parameter θ, with the following mathematical expression:

[0057]

[0058] Where η1 is the learning rate, which controls the step size of each parameter update and has a significant impact on the model's convergence speed and final performance. These are the parameters updated after one step of gradient descent. They are transitional update parameters between periods t and t+1, mainly laying the foundation for the subsequent update of the noise label model parameters w.

[0059] S6: Input the clean labels from the validation dataset into the updated image recognition model and calculate the validation set loss. Specifically, input the clean labels from the validation dataset into the updated image recognition model and calculate the validation set loss. The data input to the parameters has been updated. In the image recognition model, according to the formula Calculate the validation set loss, where the specific loss function l also adopts the cross-entropy loss function.

[0060] S7: Calculate the gradient of the validation set loss with respect to the noise label correction model parameters, and update the noise label correction model parameters.

[0061] Specifically, due to the validation set loss It includes two gradients: one for a noise label correction model and the other for an image recognition model, according to the formula. Calculate the gradient of the validation set loss with respect to the noise label corrected model parameters w.

[0062] Furthermore, a one-step gradient descent approach is used to update the noise labels and correct the model parameters w, as expressed mathematically below:

[0063]

[0064] Where w (t+1) η1 represents the updated parameters, and η2 represents the learning rate. Its function is similar to that of the learning rate η1 when training an image recognition model. It is used to adjust the step size of parameter updates to ensure that the noise label correction model can be optimized within a reasonable range, so as to improve the ability to correct noise labels.

[0065] S8: Update image recognition model parameters

[0066] Specifically, because the parameters w and θ have a nested relationship, when the noise label correction model parameter w is updated to w (t+1) After that, the parameters of the image recognition model need to be adjusted. This will be updated again. The mathematical formula is expressed as follows:

[0067]

[0068] S9: Performing image recognition tasks using updated image recognition models

[0069] Specifically, the updated image recognition model is used as the application image recognition model for subsequent image recognition tasks. Then, the data in the test dataset is input into the application image recognition model, and the image recognition results are output.

[0070] Through the above complete iterative correction and training steps, the image recognition model and the noise label correction model can be continuously optimized, thereby improving the training accuracy of the image recognition model, enabling it to better adapt to data with noisy labels, and improving the model's performance and reliability in image recognition tasks.

[0071] The above content is Embodiment 1 of this application. Embodiment 2 of this application is provided below. The only difference between Embodiment 2 of this application and Embodiment 1 of this application is that the dataset selected when creating the dataset in S1 is different.

[0072] S1: Create the dataset. The dataset includes a training dataset, a test dataset, and a validation dataset. The training dataset includes noisy labels, the validation dataset includes only clean labels, and the test dataset is the dataset of images to be recognized.

[0073] Specifically, the CIFAR-100 dataset can be downloaded from the official website. It contains 100 categories and 50,000 images. 40,000 images are used as the training dataset, 10,000 images are used as the test dataset, and 20 images are randomly selected from each class in the training dataset as the validation dataset. This ensures that the validation dataset is representative of all classes in the training dataset, but the sample size is much smaller than that of the training dataset.

[0074] Figure 4 This image shows a comparison of accuracy calculation results on the noisy CIFAR-10 and CIFAR-100 datasets, provided as embodiments of the present invention. Accuracy measures whether the results returned by the base classification model are consistent with the corresponding labels. On the CIFAR-10 and CIFAR-100 image datasets, with symmetric noise of 20%, 40%, 60%, and 80% and asymmetric noise of 20% and 40%, respectively, the method of the present invention achieved the highest image recognition accuracy. Especially on large-scale class datasets like CIFAR-100, the method of the present invention maintains its advantage, effectively classifying even with a high proportion of noise. For example, with a symmetric noise proportion of 60%, a high accuracy of 60.23% is still achieved.

[0075] On the CIFA-10 and CIFAR-100 datasets, the method of this invention demonstrates superior performance. It achieves optimal results for various levels of noise interference, whether symmetrical noise with proportions of 20%, 40%, 60%, and 80%, or asymmetrical noise with proportions of 20% and 40%. Especially on datasets like CIFAR-100 containing a large number of classes, the method maintains a significant advantage even with a high proportion of noise. For example, with symmetrical noise at 60%, the method achieves a high accuracy of 60.23%, strongly demonstrating its adaptability to complex noisy environments.

[0076] Figure 5 This is a heatmap provided in this embodiment of the invention for calculating label correction accuracy on the CIFAR-10 dataset containing 40% symmetrical noise. Label correction accuracy measures whether the pseudo-labels output by the noise label correction model are consistent with the correct labels. Figure 5As can be intuitively seen, the noise label correction model in this invention can effectively correct noise labels. Through in-depth analysis of the diagonal elements of the probability distribution matrix, it was found that the corrected label probability is significantly improved in most cases, generally exceeding 90%. This achievement not only fully demonstrates the high efficiency of the proposed noise label correction model in identifying and correcting erroneous labels, but also further highlights its excellent robustness in complex noise environments, providing reliable technical support for solving noise data problems in practical applications.

Claims

1. A method for noise-resistant label image recognition based on meta-learning, characterized in that, Includes the following steps: S1. Construct the validation dataset, training dataset, and image dataset to be recognized; S2. Input a batch of training data from the training dataset into the image recognition model and output the predicted labels; S3. Input the predicted labels and the initial labels from the training data into the noise label correction model to generate candidate correction labels; S4. A dynamic weighting strategy is used to weight and fuse the predicted label, the initial label of the training data, and the candidate correction label to generate an aggregated correction label, specifically including: The predicted label, the initial label of the training data, and the candidate corrected label are input into the noise label correction model. After being combined by weights, the aggregated corrected label is obtained, and the mathematical formula is expressed as follows: ； in Indicates aggregated correction label, Indicates the training period. It is a weight value. Indicates the initial labels of the training data. Indicates the predicted label; These are candidate correction labels, represented by the aggregated correction labels from the previous training cycle; S5. Using aggregated correction labels and validation datasets, the parameters of the image recognition model and noise label correction model are optimized through a meta-learning parameter optimization strategy. S6. Select the next batch of training data from the training dataset, and repeat steps S2 to S5 until the parameters of the two models converge. Use the image recognition model with converged parameters as the application image recognition model, input the image dataset to be recognized into the application image recognition model, and output the final image recognition result.

2. The method for noise-resistant label image recognition based on meta-learning according to claim 1, characterized in that, In step S1, the validation dataset contains only clean labels, while the training dataset contains both noisy and clean labels.

3. The method for noise-resistant label image recognition based on meta-learning according to claim 2, characterized in that, In step S2, the image recognition model adopts a deep convolutional neural network architecture, extracts image features through multi-layer convolution operations, and maps them to the label space through a fully connected layer to generate recognition results.

4. The method for noise-resistant label image recognition based on meta-learning according to claim 1, characterized in that, In step S5, the meta-learning parameter optimization strategy includes inner-layer optimization and outer-layer optimization.

5. The method for noise-resistant label image recognition based on meta-learning according to claim 4, characterized in that, In step S5, the inner layer optimization updates the image recognition model parameters by aggregating and correcting labels, and the outer layer optimization updates the noise label correction model parameters by updating the validation dataset.

6. The method for noise-resistant label image recognition based on meta-learning according to claim 5, characterized in that, In step S5, the inner layer optimization updates the image recognition model parameters by aggregating and correcting labels, and the outer layer optimization updates the noise labels and corrects the model parameters by updating the validation dataset. Specifically, this includes: First, a batch of training data from the training dataset is input into the image recognition model, and its parameters are optimized using the stochastic gradient descent method. The mathematical formula is as follows: ； in, This represents the intermediate value after the image recognition model parameters have been optimized. This represents the parameters of the image recognition model before optimization. Indicates the learning rate. Indicates the number of training samples. Represents the loss function. Represents the parameters of the loss function Perform differentiation. Represents an image recognition model. This represents a noise label correction model. This represents the data in the training dataset. This represents the label of the corresponding training data. This represents the parameters of the noise label correction model before optimization; Subsequently, a batch of validation samples from the validation dataset is input into the noise label correction model, and its parameters are optimized using the stochastic gradient descent method. The mathematical formula is as follows: ； in This represents the optimized noise label correction model parameters. Indicates the learning rate. It is the number of validation samples. Represents the loss function. Represents the parameters of the loss function Perform differentiation. This indicates that the data in the validation dataset is being validated. This represents the label corresponding to the verification data; Then use the parameters Optimize parameters The mathematical formula is as follows: ； in This represents the parameters of the optimized image recognition model.

Images