Noisy label image recognition method based on active forgetting and noise suppression

Abstract

This invention relates to the fields of artificial intelligence and machine learning, and discloses a method for noisy labeled image recognition based on active forgetting and noise suppression. The method first uses a Gaussian mixture model to dynamically partition training samples into clean and noisy subsets. For clean samples, basic supervised learning is maintained; for noisy samples, a dual regularization mechanism is introduced in parallel: one is an active forgetting module based on negative cross-entropy loss, which erases erroneous knowledge internalized by the model; the other is a noise suppression module based on complementary labels, which prevents further assimilation of noisy data by the model through a negative learning strategy. Both work together as plug-and-play regularization terms to construct a unified end-to-end optimization objective. This method effectively overcomes the limitations of traditional paradigms, exhibiting superior robustness and generalization performance on both synthetic and real noise benchmarks, providing a new technical direction for building highly reliable deep learning models.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and machine learning technology, and specifically to a method for recognizing noisy labeled images based on active forgetting and noise suppression. Background Technology

[0002] With the rapid development of deep learning technology, deep neural networks have achieved remarkable results in computer vision tasks such as image classification and object detection. These achievements largely rely on large-scale and accurately labeled datasets. However, in practical applications, obtaining high-quality labeled data is often time-consuming, labor-intensive, and costly. Researchers and industry increasingly use web crawlers, crowdsourced labeling, or automated labeling pipelines to construct datasets. This approach inevitably introduces a large number of noisy labels, meaning that the true category of the image content is inconsistent with the given label. Because modern deep neural networks typically have a large number of parameters and powerful memory capabilities, they can not only learn effective patterns in the data but also easily remember noisy samples with incorrect labels. This overfitting to noisy data can severely distort the optimization direction of the model, leading to a significant decrease in the model's generalization performance on the test set. Most existing noisy label learning methods follow a paradigm that relies on clean samples, mainly employing strategies such as sample selection, label correction, and sample reweighting. For example, they use the small loss principle to select reliable samples, use probability transition matrices to correct incorrect labels, or reduce the weight of noisy samples through meta-learning, attempting to mitigate the impact of label noise.

[0003] However, the aforementioned existing technologies still have significant limitations and technical bottlenecks in practical applications. On the one hand, in the absence of prior knowledge, existing screening mechanisms struggle to accurately distinguish between clean and noisy samples, inevitably misclassifying some noisy samples as clean and using them for training, thus limiting model performance. More critically, existing research indicates that models not only learn from clean samples in the early stages of training but also rapidly fit some noisy samples, leading to the internalization of erroneous knowledge. Relying solely on subsequent sample screening lacks a mechanism to proactively forget the erroneous knowledge stored in the model parameters. Furthermore, as training progresses, the model tends to overfit to those difficult-to-distinguish noisy samples, and relying solely on positive supervision signals is insufficient to effectively suppress this assimilation of incorrect labels. Therefore, there is an urgent need for an image recognition model training method that can overcome the limitations of the traditional clean-sample-dependent paradigm, proactively erasing erroneous knowledge accumulated in the early stages of model training while effectively suppressing subsequent noise overfitting. Summary of the Invention

[0004] To address the aforementioned technical shortcomings, the present invention aims to provide a noisy label image recognition method based on active forgetting and noise suppression. By constructing a dual-stream regularization framework that integrates active forgetting and noise suppression mechanisms, the method solves the problems of early memory and late overfitting of noisy labels in deep models.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for recognizing noisy labeled images based on active forgetting and noise suppression, the method comprising: Step S1: Using a noisy training dataset, train the deep neural network model using standard cross-entropy loss until the number of training iterations of the deep neural network model reaches the preset warm-up iteration threshold. The deep neural network model is an image recognition model; Step S2: After the number of training iterations exceeds the warm-up iteration threshold, perform the following operations: Step S2.1: For the current training batch, based on the standard cross-entropy loss of each sample in the noisy training dataset, the samples are dynamically divided into a clean sample subset and a noisy sample subset using a Gaussian mixture model. Step S2.2: For a clean subset of samples, calculate the basic supervised loss to maintain the deep neural network model's learning of the correct category features; Step S2.3: For the noisy sample subset, perform active forgetting and noise suppression operations in parallel; The active forgetting operation calculates negative cross-entropy loss, minimizing the prediction probability of the deep neural network model for a given label on a noisy sample. The noise suppression operation randomly generates a complementary label different from its given label for each noise sample, and calculates the negative learning loss based on the complementary label. Step S2.4: Weightedly fuse the basic supervision loss, active forgetting loss, and noise suppression loss to construct a joint loss function, and optimize the model parameters end-to-end by minimizing this joint loss function.

[0006] Preferably, in one possible implementation of the first aspect, in step S1, the deep neural network model is trained using standard cross-entropy loss, specifically as follows: For the noisy training dataset, the first There are samples, and the input image is... Given the label Standard cross-entropy loss The calculation method is as follows:

[0007] in, This represents the total number of categories in the classification task. The normalization coefficient is... Indicates a given label The Middle The element values ​​of each category, Indicates the current model parameters The output sample of the lower network belongs to the first The predicted probability value of the class.

[0008] Preferably, in one possible implementation of the first aspect, in step S2.1, the samples are dynamically divided into a clean sample subset and a noisy sample subset using a Gaussian mixture model, specifically as follows: Based on the standard cross-entropy loss values ​​of all samples in the current training batch, a binary Gaussian mixture model is fitted using the expectation-maximization algorithm to characterize the loss distribution. The binary Gaussian mixture model contains two Gaussian distribution components, and the Gaussian distribution component with the smaller mean is the small loss component. Based on the posterior probability that a sample belongs to a small loss component, the current training batch of data is dynamically divided into a clean sample subset. and noise sample subset .

[0009] Preferably, in one possible implementation of the first aspect, in step S2.2, calculating the basic supervision loss specifically involves: For samples in the clean sample subset, calculate their basic supervised loss. The expression is:

[0010] in, This indicates the total number of samples included in the clean sample subset selected in the current training batch.

[0011] Preferably, in one possible implementation of the first aspect, in step S2.3, calculating the negative cross-entropy loss specifically involves: For samples in the noisy subset, calculate their negative cross-entropy loss. The expression is:

[0012] in, For the noise sample subset, the first The negative cross-entropy loss of each sample.

[0013] Preferably, in one possible implementation of the first aspect, in step S2.3, calculating the negative learning loss based on complementary labels specifically involves: For samples in the noisy subset, calculate their negative learning loss. The expression is:

[0014] in, For the noise sample subset, the first The negative learning loss for each sample. Indicates the first The complementary label generated from the noise sample is the first one. The element value of the category, when the complementary label is the . The value of this element is 1 if it is of class 1, otherwise it is 0.

[0015] Preferably, in one possible implementation of the first aspect, the rule for generating the complementary tag is: For a noisy sample in a subset of noisy samples, randomly select a category different from the given label of the noisy sample from all categories in the current classification task as the complementary label; The complementary label is the true negative class of the noise sample.

[0016] Preferably, in one possible implementation of the first aspect, in step S2.4, the joint loss function The expression is:

[0017] in, This represents the subset of noisy samples selected in the current training batch; This represents the total number of samples in the noise sample subset, used to calculate the average value; and These are preset hyperparameters, used to adjust the relative importance of the active forgetting module and the noise suppression module in the overall optimization objective.

[0018] In a second aspect, the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the noisy label image recognition method based on active forgetting and noise suppression as described in the first aspect.

[0019] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the noisy label image recognition method based on active forgetting and noise suppression as described in the first aspect.

[0020] The beneficial effects of this invention are as follows: This invention proposes a robust learning method for noisy labeled image recognition. By constructing a dual-stream regularization framework that integrates an active forgetting mechanism and a noise suppression mechanism, it effectively solves the problems of early memory and late overfitting of noisy labels in deep models, providing a new perspective for learning with noisy labels.

[0021] This invention first identifies two key phenomena behind noise fitting: noise memory resulting from generalization learning in the early stages, and noise overfitting in the later stages. To address these challenges, two lightweight, plug-and-play modules are integrated: an active forgetting module based on machine forgetting, which erases accumulated noise through negative cross-entropy loss; and a noise suppression module based on negative learning, which uses complementary labels to prevent further noise assimilation.

[0022] Together, these two elements form a unified mechanism that proactively eliminates erroneous knowledge and suppresses future noisy fitting without introducing architectural complexity. Extensive experiments on synthetic and real-world benchmarks demonstrate that this invention continuously improves the robustness and generalization ability of the model, providing a new direction for more noise-tolerant learning frameworks. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This application provides a flowchart of a method for recognizing noisy labeled images based on active forgetting and noise suppression. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1: As Figure 1 As shown, this invention provides a method for recognizing noisy labeled images based on active forgetting and noise suppression, comprising: Step S1: Using a noisy training dataset, train the deep neural network model using standard cross-entropy loss until the number of training iterations of the deep neural network model reaches a preset warm-up iteration threshold; the deep neural network model is an image recognition model.

[0027] Specifically, the deep neural network model applicable to this invention is a general image recognition model architecture. Given that deep neural networks have a dynamic characteristic of first fitting simple general patterns of data and then gradually fitting complex noisy data during the training process, the entire training process of the deep neural network model is divided into two continuous stages in the time dimension: the early generalization learning stage and the later robust learning stage.

[0028] To fully utilize the memory effect of deep neural network models in the early stages of training, a warm-up phase is first implemented using a noisy training dataset, with a pre-warm-up iteration threshold set. During the generalization learning phase, where the number of iterations is less than this pre-warm-up threshold, a standard supervised learning strategy is used to initially optimize the deep neural network. For the [number of iterations] in the training dataset... Let there be samples, denoted as the input image. and given label The input is fed into the network to obtain the predicted probability distribution, and the difference between this predicted probability distribution and the given label is calculated. This difference is measured using the standard cross-entropy loss function, and its calculation formula is as follows:

[0029] In this formula, This represents the total number of categories for the classification task. These are the normalization coefficients; Indicates the first The first sample with a given label The element value of the nth category, when the label of the sample belongs to the nth category. The value is 1 if the condition is met, and 0 otherwise. Indicates the current model parameters Next, the network output of the first The sample belongs to the first The predicted probability value of the class. By minimizing this loss function during the generalization learning phase, the deep neural network model can quickly capture the main feature distribution in the data and establish a preliminary feature representation capability.

[0030] Step S2: After the number of training iterations exceeds the preheating iteration threshold, perform the following operations.

[0031] Step S2.1: For the current training batch, based on the standard cross-entropy loss of each sample, the samples are dynamically divided into a clean sample subset and a noisy sample subset using a Gaussian mixture model.

[0032] Specifically, after the number of training iterations exceeds the warm-up threshold and enters the robust learning phase, this invention first performs a data splitting operation to distinguish between reliable and unreliable data, providing a foundation for the subsequent denoising mechanism. In practice, the loss value of all samples in the current batch is calculated, and a binary Gaussian mixture model is fitted to represent the loss distribution using the expectation-maximization algorithm. The binary Gaussian mixture model contains two Gaussian distribution components; the Gaussian distribution component with the smaller mean is the smaller loss component. Based on the posterior probability that a sample belongs to the smaller loss component, the current batch of data is dynamically divided into a clean subset with high confidence. and a subset of noisy samples with low confidence The judgment process specifically involves setting a selection threshold. In this embodiment, the selection threshold is set to (1 - the estimated overall noise rate in the dataset). Samples with a posterior probability greater than or equal to the selection threshold are classified as a clean sample subset with high confidence, while samples with a posterior probability lower than the selection threshold are classified as a noisy sample subset with low confidence.

[0033] Step S2.2: For a clean subset of samples, calculate the basic supervised loss to maintain the model's learning of the correct category features.

[0034] Specifically, for the selected clean sample subset This invention performs basic supervised learning to maintain and reinforce the deep neural network's ability to extract features from the real class distribution. Basic supervised loss. The calculation formula is as follows:

[0035] In this formula, This indicates the total number of samples included in the clean sample subset selected in the current batch. Denotes the first clean sample in the subset. A number of image samples and their corresponding given labels. The standard cross-entropy loss function is used to optimize the model and ensure it is anchored to the correct data manifold.

[0036] Step S2.3: For a subset of noisy samples, perform active forgetting and noise suppression operations in parallel. The active forgetting operation calculates the negative cross-entropy loss to minimize the model's prediction probability for a given label of a noisy sample. The noise suppression operation randomly generates a complementary label for each noisy sample that is different from its given label and calculates the negative learning loss based on the complementary label.

[0037] Specifically, for data classified as belonging to the noisy sample subset, this invention no longer performs conventional supervised learning but instead introduces a dual regularization mechanism in parallel: first, an active forgetting module based on negative cross-entropy loss to erase erroneous knowledge already internalized by the model; second, a noise suppression module based on complementary labels to prevent further assimilation of noisy data by the model through a negative learning strategy. The core purpose of the machine-based active forgetting mechanism is to reverse the overfitting process of these mislabeled samples during the warm-up phase, actively erasing the erroneous category information already internalized in the model parameters.

[0038] For each sample in the noisy subset, calculate its negative cross-entropy loss. Unlike the standard cross-entropy loss, which attempts to maximize the model's predicted probability for a given label, the negative cross-entropy loss minimizes the model's predicted probability for a given noisy label. Mathematically, this is achieved by taking a positive value of the logarithmic probability. When this positive loss is minimized, the optimization algorithm drives the corresponding probability value of the model output to tend towards zero, thus achieving the "forgetting" of specific erroneous associations. Its calculation formula is as follows:

[0039] In this formula, This represents the total number of categories for the classification task. For the normalization coefficient used to balance the dimensions; Indicates the first In a given label of a noise sample, the first The element values ​​of each category, Indicates the current model parameters Next, the network output of the first The noise sample belongs to the first The predicted probability value of the class. By introducing this loss term, the optimization direction of the gradient is effectively reversed, causing the model parameters to be updated in the direction of reducing the confidence of the noise label, thereby specifically eliminating noise interference accumulated in the early stage while preserving the image feature representation.

[0040] To further prevent the model from refitting noisy data after active forgetting, or from generating high confidence in incorrect categories in the absence of strong supervision signals, a noise suppression mechanism based on complementary labels is introduced simultaneously with active forgetting. This mechanism utilizes a negative learning strategy to provide the model with an indirect supervision signal that "the input image does not belong to a specific category," thereby constraining the model's feature learning space without relying on potentially incorrect given labels and effectively suppressing overfitting in the later stages.

[0041] First, a complementary label is generated for each sample in the noisy sample subset. The generation rule is: from all categories in the current classification task, a category different from the given label of the sample is randomly selected as the complementary label. Since classification tasks typically have many categories, this complementary label is almost certainly the true negative class of the image. Then, a negative learning loss function is constructed using this complementary label. The goal of this loss function is to force the model to predict that the probability of the sample belonging to the complementary label category approaches zero. The calculation formula is as follows:

[0042] In this formula, This represents the total number of categories for the classification task. These are the normalization coefficients; Indicates the first The complementary labels generated from the nth sample are the nth... The element value of the category, when the complementary label is the . The value is 1 if the condition is met, otherwise it is 0. Indicates the current model parameters Below, the network outputs that the sample belongs to the first... The predicted probability value of the class. The logarithmic term in the formula. This results in a smaller loss value as the model's predicted probability for complementary categories decreases. By minimizing this loss, the model reduces its prediction confidence for complementary label categories, effectively preventing the model from blindly fitting noisy labels or arbitrary incorrect categories.

[0043] Step S2.4: Weightedly fuse the basic supervision loss, active forgetting loss, and noise suppression loss to construct a joint loss function, and optimize the model parameters end-to-end by minimizing this joint loss function.

[0044] Specifically, to integrate supervised learning, active forgetting mechanisms, and noise suppression mechanisms into a unified training framework and achieve end-to-end joint optimization of deep neural network parameters, this invention constructs a total objective loss function. This function weights and fuses the supervised loss on clean samples and the double regularization loss on noisy samples, balancing the contributions of different modules to gradient updates.

[0045] In each iteration of the robust learning phase, the total loss for the entire batch of data is calculated. This total loss consists of three parts: the first part is the average basic supervised loss on the clean sample subset; the second part is the average active forgetting loss on the noisy sample subset, multiplied by a first weighting factor; and the third part is the average noise suppression loss on the noisy sample subset, multiplied by a second weighting factor. The formula for calculating the total loss function is as follows:

[0046] In this formula, For clean sample-based supervision loss; This represents the subset of noise samples selected in the current batch; This represents the total number of samples in the subset, used to calculate the average. and These are the active forgetting loss value and the noise suppression loss value calculated for a single noise sample, respectively. and The hyperparameters are preset and used to adjust the relative importance of the active forgetting module and the noise suppression module in the overall optimization objective. The overall loss function is minimized using the stochastic gradient descent algorithm. And update the model parameters through backpropagation. This process continues until the preset maximum number of iterations is reached, ultimately yielding an image recognition model with high robustness and generalization ability.

[0047] Example 2: This invention provides a training process for a noisy labeled image recognition method based on active forgetting and noise suppression. This example uses three widely used real-world noise benchmarks: Web-Aircraft, Web-Bird, and Web-Car. These datasets are compiled from web image search engines and represent subsets of the large-scale, fine-grained WebFG-496 dataset. Unlike synthetic impaired datasets, real-world benchmarks involve complex and unknown noise patterns, making performance improvement more challenging.

[0048] For real-world noise benchmarks, this embodiment uses a ResNet-50 backbone network pre-trained on ImageNet-1K. The batch size is set to 32, the training process uses a stochastic gradient descent optimizer with a momentum of 0.9, and a cosine annealing learning rate scheduling strategy is employed to balance the hyperparameters of the total loss weights. and The default values ​​are 0.001 and 0.1, respectively.

[0049] Example 3: This invention provides a comparative experimental result of a noisy labeled image recognition method based on active forgetting and noise suppression. This example compares various learning paradigms, including sample selection, label correction, robust loss design, and sample reweighting.

[0050] The comparison methods used in this embodiment include Decopupling, Co-teaching, Co-teaching+, PENCIL, JoCoR, AFM, DivideMix, Jo-SRC, Co-LDL, UNICON, SOP, AGCE, DISC, ANL, NPN, CA2C, ACT, and SED. ACT and SED are used as the basic framework, with this invention integrated as a plug-and-play regularizer. The accuracy comparison results of the proposed method and the comparison methods on real-world noisy datasets are shown in Table 1. All results are the average of five runs with different random seeds.

[0051] Table 1. Accuracy comparison of the method of the present invention and the comparative method on real-world noisy datasets.

[0052] Integrating this invention into two benchmarks (SED and ACT) resulted in improved accuracy on all three real-world noise benchmarks. This result demonstrates the strong adaptability of the method and its practical value as a plug-and-play module.

[0053] Example 4: This example provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, performs the steps of the method involved in Example 1.

[0054] Example 5: This example provides a computer program product containing instructions, including a memory, a processor, and a program stored in the memory and executable on the processor, which, when run on a computer, causes the computer to perform the steps of the method involved in Example 1.

[0055] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for recognizing noisy labeled images based on active forgetting and noise suppression, characterized in that, The method includes: Step S1: Using a noisy training dataset, train the deep neural network model using standard cross-entropy loss until the number of training iterations of the deep neural network model reaches the preset warm-up iteration threshold. The deep neural network model is an image recognition model; Step S2: After the number of training iterations exceeds the warm-up iteration threshold, perform the following operations: Step S2.1: For the current training batch, based on the standard cross-entropy loss of each sample in the noisy training dataset, the samples are dynamically divided into a clean sample subset and a noisy sample subset using a Gaussian mixture model. Step S2.2: For a clean subset of samples, calculate the basic supervised loss to maintain the deep neural network model's learning of the correct category features; Step S2.3: For the noisy sample subset, perform active forgetting and noise suppression operations in parallel; The active forgetting operation calculates negative cross-entropy loss, minimizing the prediction probability of the deep neural network model for a given label on a noisy sample. The noise suppression operation randomly generates a complementary label different from its given label for each noise sample, and calculates the negative learning loss based on the complementary label. Step S2.4: Weightedly fuse the basic supervision loss, active forgetting loss, and noise suppression loss to construct a joint loss function, and optimize the model parameters end-to-end by minimizing this joint loss function.

2. The noisy label image recognition method based on active forgetting and noise suppression as described in claim 1, characterized in that, In step S1, the deep neural network model is trained using standard cross-entropy loss, specifically as follows: For the noisy training dataset, the first There are samples, and the input image is... Given the label Standard cross-entropy loss The calculation method is as follows: in, This represents the total number of categories in the classification task. The normalization coefficient is... Indicates a given label The Middle The element values ​​of each category, Indicates the current model parameters The output sample of the lower network belongs to the first The predicted probability value of the class.

3. The noisy label image recognition method based on active forgetting and noise suppression as described in claim 2, characterized in that, In step S2.1, the samples are dynamically divided into a clean sample subset and a noisy sample subset using a Gaussian mixture model, specifically as follows: Based on the standard cross-entropy loss values ​​of all samples in the current training batch, a binary Gaussian mixture model is fitted using the expectation-maximization algorithm to characterize the loss distribution. The binary Gaussian mixture model contains two Gaussian distribution components, and the Gaussian distribution component with the smaller mean is the small loss component. Based on the posterior probability that a sample belongs to a small loss component, the current training batch of data is dynamically divided into a clean sample subset. and noise sample subset .

4. The noisy label image recognition method based on active forgetting and noise suppression as described in claim 3, characterized in that, In step S2.2, the basic supervision loss is calculated, specifically as follows: For samples in the clean sample subset, calculate their basic supervised loss. The expression is: in, This indicates the total number of samples included in the clean sample subset selected in the current training batch.

5. The noisy label image recognition method based on active forgetting and noise suppression as described in claim 4, characterized in that, In step S2.3, the negative cross-entropy loss is calculated, specifically as follows: For samples in the noisy subset, calculate their negative cross-entropy loss. The expression is: in, For the noise sample subset, the first The negative cross-entropy loss of each sample.

6. The noisy label image recognition method based on active forgetting and noise suppression as described in claim 5, characterized in that, In step S2.3, the negative learning loss based on complementary labels is calculated as follows: For samples in the noisy subset, calculate their negative learning loss. The expression is: in, For the noise sample subset, the first The negative learning loss for each sample. Indicates the first The complementary label generated from the noise sample is the first one. The element value of the category, when the complementary label is the . The value of this element is 1 if it is of class 1, otherwise it is 0.

7. The noisy label image recognition method based on active forgetting and noise suppression as described in claim 6, characterized in that, The rules for generating the complementary tags are as follows: For a noisy sample in a subset of noisy samples, randomly select a category different from the given label of the noisy sample from all categories in the current classification task as the complementary label; The complementary label is the true negative class of the noise sample.

8. The noisy label image recognition method based on active forgetting and noise suppression as described in claim 6, characterized in that, In step S2.4, the joint loss function The expression is: in, This represents the subset of noisy samples selected in the current training batch; This represents the total number of samples in the noise sample subset, used to calculate the average value; and These are preset hyperparameters, used to adjust the relative importance of the active forgetting module and the noise suppression module in the overall optimization objective.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by a processor, the program implements the steps in the noisy label image recognition method based on active forgetting and noise suppression as described in any one of claims 1-8.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the noisy label image recognition method based on active forgetting and noise suppression as described in any one of claims 1-8.

Images