Anti-noise tag image recognition method based on meta-learning

Abstract

The invention discloses an anti-noise label image recognition method based on meta-learning. The method comprises the following steps: (1) data set construction: constructing a verification data set, a training data set and a to-be-recognized image data set; (2) noise label correction: innovatively fusing a data initial label, a candidate correction label and an image identification model prediction label, and generating an aggregation correction label to replace an original noise label by integrating information of the three labels through a weight coefficient; and then updating the noise label correction model by means of a gradient descent algorithm so as to reduce the loss of the verification data set to the greatest extent. And (3) optimization of meta-learning parameters: constructing a joint optimization target containing image recognition model parameters and noise label correction model parameters, and realizing collaborative gradient updating of two parameter spaces through a dual optimization algorithm. The method can be widely applied to image recognition data sets, in particular to recognition of image data sets which are disclosed on the network and contain noise tags.

Description

Technical Field

[0001] The present 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 Art

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

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

[0004] Currently, methods for addressing this problem are mainly categorized into sample selection and label correction. Sample selection methods address this challenge by applying importance weights to each sample based on its loss value. Some work has made the weighting scheme more adaptable by using a small amount of validation data to guide the network training process. These methods aim to discard suspicious noise samples during training. However, even if the discarded samples are mostly noise, they often contain useful information that can improve the accuracy and robustness of the network, especially in situations with high noise ratios.

[0005] Label correction methods alleviate this problem by attempting to find and correct noisy labels to their underlying true labels. Some work attempts to estimate the noise transfer matrix, which is the probabilistic mapping from true labels to noisy labels, and then use the estimated matrix to correct the corrupted samples. However, as the number of categories increases, the matrix size grows exponentially, making it complex and intractable to correct noisy labels on large-scale datasets. In addition, these methods assume that the noise is class-dependent, which does not work well for more complex noise (such as feature-dependent noise).

[0006] Therefore, studying how to robustly train image recognition models in the presence of noisy labels in the training data not only has important academic research value, but also has important significance for promoting the practical application of deep learning in more fields. Summary of the Invention

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

[0008] The object of the present 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 a verification dataset, a training dataset, and a dataset of images to be identified;

[0010] S2. Input the small batch training data in the training dataset into the image recognition model and output the predicted label;

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

[0012] S4, using a dynamic weight strategy to perform weighted fusion on the predicted label, the initial label of the training data, and the candidate correction label to generate an aggregated correction label;

[0013] S5. Constructing a meta-learning parameter optimization strategy, wherein the meta-learning parameter optimization strategy includes inner layer optimization and outer layer optimization;

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

[0015] Furthermore, the step S1 specifically includes: selecting a small amount of clean data from the noisy data set to form a validation data set, and the remaining data as a training data set. The mathematical formula is as follows:

[0016]

[0017] in represents the training data set, n represents the number of samples in the training data set, x i Represents the i-th training data in the training data set, y i Represents the label corresponding to the i-th training data, Represents the validation data set, m represents the number of samples in the validation data set (its number is much smaller than the number of samples in the training data set), Represents the i-th verification data in the verification data set, Denotes 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] Furthermore, the step S4 specifically includes: inputting the noise label, the predicted label and the candidate correction label into the noise label correction model, and outputting the aggregated correction label, which is expressed by the following mathematical formula:

[0019]

[0020] in represents the aggregated corrected label, t represents the training cycle, g represents the noise label correction model, y represents the initial label of the training data (which may be a noise label), represents the predicted label output by the image recognition model, represents the candidate correction label (the aggregated correction label of the previous training cycle, which is used to alleviate the negative impact of significant fluctuations in the prediction results of the image recognition model during the training process and improve the accuracy of the correction label), w (t) represents the parameters of the noise label correction model.

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

[0022]

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

[0024] Furthermore, step S5 specifically includes: using the aggregated correction labels generated by the noise label correction model and the verification data set, and jointly training the image recognition model and the noise label correction model through a meta-learning parameter optimization strategy, which is expressed in the following mathematical formula:

[0025]

[0026] where θ * represents the optimal parameters of the image recognition model, θ * (w) indicates that the parameters w and θ are nested, θ represents the parameters of the image recognition model, and w represents the parameters of the noise correction model. represents the loss of the training dataset, w * represents the optimal parameters of the noise label correction model, represents the validation dataset loss, f represents the noise label correction model, and l represents the loss function, such as the cross entropy loss function. When the noise label correction model parameters reach the optimal value, that is, w = w * , which tends to correct noisy labels into 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 θ * (w), that is, every time parameter w is updated in gradient descent, parameter θ must also be updated synchronously. Furthermore, the computational relationship between the two parameters at each step must be stored until both are updated to the optimal value. This nested parameter optimization consumes a lot of computing resources and is almost impossible to implement in practice. Therefore, this method utilizes local gradient information, using the parameters obtained after a single step of gradient descent to approximate the optimal parameters for update. The specific steps are as follows:

[0028] First, the small batch training samples in the training dataset are recorded as Training dataset loss It can be expressed as: We can further derive the update formula of parameter θ with respect to parameter w in single-step gradient descent:

[0029]

[0030] in Represents the intermediate result of parameter θ update, θ (t) represents the image recognition model parameters at training cycle t, η1 represents the learning rate, Then, we use the small batch validation samples in the validation dataset to Update the parameter w by a single step of gradient descent:

[0031]

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

[0033]

[0034] Among them, θ (t+1) 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 traditional predefined generation rules and can automatically map noisy labels to corrected labels. This approach greatly improves the flexibility of the noisy label correction process, making it better adaptable to complex real-world image recognition datasets with diverse types and varying noise levels.

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

[0038] 3. The present invention does not rely on a specific model structure and has strong versatility. It can be easily applied to mainstream image recognition models such as CNN and Transformer. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0040] Figure 2 A flowchart of the steps provided in an embodiment of the present invention;

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

[0042] Figure 4 A comparison chart of the recognition accuracy calculated on the noisy CIFAR-10 and CIFAR-100 datasets provided by an embodiment of the present invention;

[0043] Figure 5 A heat map showing the accuracy of noise label correction calculated on the CIFAR-10 dataset containing 40% symmetric noise, provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The following will be combined with the accompanying drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the embodiments described are only part of the embodiments of the present application, not all of the embodiments.

[0045] This application embodiment provides a noise label correction method based on meta-learning. The specific implementation process is as follows: Figure 2 As shown, it covers steps S1 to S9. It should be noted that Figure 2 What is shown is only one execution order presented in the embodiment of the present application, and is not the only execution order of the method. In the case of achieving the final result, each step can be executed in parallel or in reverse. Figure 3 The noise label correction model structure shown and the image recognition model structure described in the article are general situations presented in the embodiments of this application, and are not unique structures. In actual applications, they can be flexibly adjusted and optimized according to specific needs and scenarios.

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

[0047] Specifically, the CIFAR-10 dataset can be downloaded from the official website. It contains 60,000 images in 10 categories. 50,000 images are used as the training dataset, and 10,000 images are used as the test dataset. 100 images are randomly selected from each class in the training dataset as the validation dataset, ensuring that the validation dataset is representative of all classes in the training dataset. However, the sample size is much smaller than that of the training dataset. The training dataset is represented as follows: Consists of n samples, for each image x i , whose corresponding label is y i Two types of noise are selected to add noise to the labels of the training dataset, and the noise ratio is set to p: (1) Symmetric noise: A clean label with probability Convert to other tags and The probability of keeping the original label is , where k is the number of labels. (2) Asymmetric noise: A clean sample label is corrupted into other labels 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. It mainly helps the noise label correction model learn the correct semantics.

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

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

[0050] Specifically, the image recognition model consists of ResNet (deep convolutional neural network) and is pre-trained on Imagenet (image dataset), which is expressed as: Where x represents the data input to the model, θ is the parameter of the image recognition model, is the output prediction result. After completing fine-tuning and pre-training, the training data is input into the image recognition model, according to the formula Compute the predicted labels for the samples.

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

[0052] Specifically, the predicted label obtained in step S2 Initial label y i and candidate correction labels Input into the noise label correction model, and combine them through weights to obtain the aggregated corrected label It should be noted that when y=1, i.e. the first training cycle, due to the lack of candidate correction labels, we set That is, the predicted labels of the image recognition model are used to replace the candidate corrected labels for 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, according to the formula Calculate the training set loss. The specific loss function l is the cross-entropy loss function. The cross-entropy loss function can converge to the optimal solution more quickly during backpropagation, effectively avoiding the vanishing gradient problem, thereby ensuring the stability and efficiency of model training.

[0054] Furthermore, due to the loss of training set Covering two different gradients of 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 value plays a key role in the subsequent update of the image recognition model parameters. It reflects the direction of the model's error on the current training set, allowing the model to be optimized in the direction of reducing the error.

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

[0056] Specifically, since the parameters w and θ have a nested relationship θ * (w), so when updating the parameter θ, it is necessary to fully preserve the computational relationship between the two at each step to ensure the consistency and accuracy of the computational logic of the entire model. In order to reduce the above storage and computing resources, a one-step gradient descent is used to update the image recognition model parameter θ. The mathematical expression is as follows:

[0057]

[0058] Among them, η1 is the learning rate, which controls the step size of each parameter update and has an important impact on the convergence speed and final performance of the model. It is the parameter updated after one step of gradient descent. It is the transition update parameter between cycles t and t+1, which mainly lays the foundation for the update of the subsequent noise label model parameter w.

[0059] S6: Input the clean labels in the validation dataset into the updated image recognition model and calculate the validation set loss. Specifically, the validation dataset The data input to parameter has been updated to In the image recognition model, according to the formula Calculate the validation set loss, where the specific loss function l also uses 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: the noise label correction model and the image recognition model. According to the formula Calculate the gradient of the validation set loss with respect to the noise label correction model parameters w.

[0062] Furthermore, a one-step gradient descent is used to update the noise label correction model parameter w, which is expressed as follows:

[0063]

[0064] where w (t+1) Represents the updated parameters, η2 represents the learning rate, which is similar to the learning rate η1 when training the image recognition model. It is used to adjust the step size of the parameter update to ensure that the noise label correction model can be optimized within a reasonable range to improve the correction ability of the noise label.

[0065] S8: Update image recognition model parameters

[0066] Specifically, since the parameters w and θ are nested, when the noise label correction model parameter w is updated to w (t+1) After that, the image recognition model parameters need to be adjusted. Update again. The mathematical formula is as follows:

[0067]

[0068] S9: Use the updated image recognition model to perform image recognition tasks

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

[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 containing noisy labels, and improving the performance and reliability of the model in image recognition tasks.

[0071] The above content is Example 1 of the present application. Example 2 of the present application is provided below. The difference between Example 2 of the present application and Example 1 of the present application is only that the data set selected when establishing the data set in S1 is different.

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

[0073] Specifically, the CIFAR-100 dataset, which contains 50,000 images across 100 categories, can be downloaded from the official website. 40,000 images are used for training, 10,000 for testing, 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 the training dataset.

[0074] Figure 4 The embodiment of the present invention provides a comparative result diagram of the accuracy calculated on the noisy CIFAR-10 and CIFAR-100 datasets. The accuracy measures whether the results returned by the base classification model are consistent with the corresponding labels. On the two image datasets, CIFAR-10 and CIFAR-100, with 20%, 40%, 60%, and 80% symmetrical noise and 20% and 40% asymmetric noise respectively, the method of the present invention achieved the highest image recognition accuracy. In particular, on large-scale category datasets such as CIFAR-100, the method of the present invention can still maintain its advantages and can effectively perform classification even when the noise ratio is high. For example, when the symmetrical noise ratio is 60%, a high accuracy of 60.23% can still be achieved.

[0075] On the CIFA-10 and CIFAR-100 datasets, the method of the present invention demonstrated excellent performance. For different degrees of noise interference, whether it is symmetrical noise at ratios of 20%, 40%, 60%, and 80%, or asymmetric noise at ratios of 20% and 40%, the method of the present invention achieved the best results. Especially on a dataset such as CIFAR-100 that contains a large number of categories, the method of the present invention can still maintain a significant advantage even in the face of a higher proportion of noise. Taking the case of symmetrical noise at a ratio of 60% as an example, the method of the present invention achieved a high accuracy rate of 60.23%, which strongly demonstrated its strong adaptability in complex noise environments.

[0076] Figure 5 The heat map of the label correction accuracy calculated on the CIFAR-10 dataset with 40% symmetrical noise provided by the embodiment of the present invention is a measure of whether the pseudo labels output by the noise label correction model are consistent with the correct labels. Figure 5It can be intuitively seen that the noise label correction model in the method of the present invention can effectively correct noisy labels. Through in-depth analysis of the diagonal elements of the probability distribution matrix, the study found that the probability of the corrected labels is greatly improved in most cases, generally exceeding 90%. This result not only fully confirms the high efficiency of the acoustic label correction model proposed in this invention in identifying and correcting erroneous labels, but also further highlights its excellent robustness in complex noisy environments, providing reliable technical support for solving the problem of noisy data in practical applications.

Claims

1. A method for noise-resistant label image recognition based on meta-learning, characterized in that: The following steps are involved: S1. Construct a verification dataset, a training dataset, and a dataset of images to be identified; S2: Input a batch of training data from the training dataset into the image recognition model and output the predicted label; S3. Input the predicted label and the initial label of the training data into the noise label correction model to generate candidate correction labels; S4, using a dynamic weight strategy to perform weighted fusion on the predicted label, the initial label of the training data, and the candidate correction label to generate an aggregated correction label; S5. Optimize the parameters of the image recognition model and the noise label correction model using the aggregated correction labels and the validation dataset through a meta-learning parameter optimization strategy. S6. Select the next batch of training data from the training data set, repeat steps S2 to S5 until the two model parameters converge, use the image recognition model with converged parameters as the application image recognition model, input the image data set to be recognized into the application image recognition model, and output the final image recognition result.

2. The method for anti-noise label image recognition based on meta-learning according to claim 1, characterized in that: In step S1, the validation dataset contains only clean labels, and the training dataset contains noisy labels 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 anti-noise label image recognition based on meta-learning according to claim 3, characterized in that: In step S4, a dynamic weight strategy is used to perform weighted fusion on 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 correction label are input into the noise label correction model. After combining them through weights, the aggregated correction label is obtained. The mathematical formula is as follows: in represents the aggregated corrected label, t represents the training cycle, α is the weight value, and y represents the initial label of the training data. represents the predicted label; is a candidate correction label, represented by the aggregated correction label of the previous training cycle.

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

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 the correction labels, and the outer layer optimization updates the noise label correction model parameters by verifying the dataset.

7. The method for noise-resistant label image recognition based on meta-learning according to claim 6, characterized in that: In step S5, the inner optimization updates the image recognition model parameters by aggregating the corrected labels, and the outer optimization updates the noise label correction model parameters by using the validation dataset, specifically including: 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, Represents the intermediate value after the image recognition model parameters are optimized, θ (t) represents the image recognition model parameters before optimization, η1 represents the learning rate, n represents the number of training samples, l represents the loss function, represents the derivation of the loss function parameter θ, f represents the image recognition model, g represents the noise label correction model, x i Represents the data in the training dataset, y i Indicates the label of the corresponding training data, w (t) represents the noise label correction model parameters before optimization; Subsequently, a batch of validation samples in the validation dataset is input into the noise label correction model, and its parameters are optimized according to the stochastic gradient descent method. The mathematical formula is as follows: where w (t+1) represents the optimized noise label correction model parameters, η2 represents the learning rate, m is the number of verification samples, l represents the loss function, Indicates the derivation of the loss function parameter w, Represents the data in the validation dataset, Indicates the label of the corresponding verification data; Then use the parameter w (t+1) To optimize the parameter θ (t) , the mathematical formula is as follows: where θ (t+1) Represents the optimized image recognition model parameters.

Images