A method and apparatus for training a neural network model with noisy multi-label data

Abstract

This invention relates to a method and apparatus for training a neural network model on noisy multi-label data. The method includes the following steps: selecting a clean set of samples for each category as a meta-dataset using a sample selection algorithm, and estimating the category-dependent label noise transition matrix; initializing some parameters in the instance feature-dependent label noise transition matrix network using the category-dependent label noise transition matrix; transforming the learning problem into a two-layer optimization problem based on statistically consistent label noise learning loss, and simultaneously learning the instance feature-dependent label noise transition matrix network parameters, data imbalance parameters, and multi-label classification neural network parameters using a meta-learning algorithm. This invention innovatively utilizes a meta-learning algorithm in a data-driven manner to unify the learning of instance feature-dependent label noise transition matrix network parameters, data imbalance parameters, and multi-label classification neural network parameters within a single framework.

Description

Technical Field

[0001] This invention belongs to the Internet field, and specifically relates to a method and apparatus for training a neural network model with noisy multi-label data. Background Technology

[0002] In recent years, deep learning has achieved remarkable results in fields such as image recognition, object detection, visual tracking, and text matching. This is mainly attributed to the powerful nonlinear mapping capabilities of deep neural networks, which ensure a high degree of separability in the feature space representation of data. However, with the development of deep learning, the amount of data required to train high-precision models has exploded as the required accuracy increases. For example, the ImageNet image classification dataset contains 14,197,122 images, and the CLIP cross-modal pre-trained model used 400 million text-image pairs for training. As the scale of datasets grows larger, the manual and economic costs required for accurate annotation of these large-scale datasets increase exponentially, especially in fields such as medical image processing and financial risk control. For instance, in medical image processing, accurate annotation of a lesion image requires diagnosis by one or even multiple experts, greatly increasing the difficulty of annotating large datasets; in international credit card theft scenarios in financial risk control, it typically takes six months to receive case labels.

[0003] Against this backdrop, some simple annotation methods have emerged, such as crowdsourcing methods that divide a large dataset into several smaller datasets and distribute them to different annotators, and automatic annotation methods that utilize search engines, web links, or video tags. However, while these methods obtain a large amount of labeled data at a relatively low cost, they also introduce irreversible label noise into the constructed datasets. For example, in crowdsourcing methods, different annotators have different abilities to identify data and different biases towards different categories, which can lead to different errors in labeling some data.

[0004] The following are some of the main methods for addressing the learning problem under noisy labeled samples.

[0005] One class of methods focuses on statistical consistency: by designing a loss function that allows a classifier learned from noisy data to asymptotically converge to the optimal classifier defined on the clean domain. The noise transition matrix, representing the probability of a clean label being converted to a noisy label, is used to construct a range of statistical consistency algorithms. Specifically, it has been used to modify loss functions to build label-noisy learning algorithms with risk consistency, and to constrain the hypothesis space to build label-noisy learning algorithms with classifier consistency.

[0006] The second category of methods lacks statistical consistency: these employ heuristic designs to reduce the side effects of label noise, such as extracting reliable examples, correcting labels, and adding implicit or explicit regularization. Currently effective methods for extracting reliable examples include, but are not limited to, the following: distillation, sample sieving, Gaussian mixture model loss distribution modeling, confidence-based sample sets, small-loss-based methods, and some early stopping techniques.

[0007] The third type of approach utilizes a small amount of clean data: it resists the influence of label noise by leveraging the distribution information of a small number of clean labels. Most work on label noise learning assumes that all training data labels may be incorrect. However, there are usually some reliable examples available to create validation and test sets. By assuming that a subset of the training data is reliable, the approach of utilizing a small amount of clean data changes the assumption that all training data may be corrupted, and it has been shown that having a certain amount of reliable training data can significantly improve robustness. These methods mostly employ meta-learning forms to constrain the hypothesis space.

[0008] Although researchers have proposed numerous methods for label-noise deep learning, most of these methods focus on simple multi-class noisy learning scenarios. However, due to differences in label formats and learning methods, real-world applications often involve diverse scenario characteristics, making these methods difficult to transfer easily. Specifically, in noisy multi-label learning scenarios, each sample contains multiple target categories, and the label for each category may contain noise. Few methods in multi-label learning address the consequences of label noise. Therefore, proposing a method for training neural network models on noisy multi-label data is essential. Summary of the Invention

[0009] This invention provides a method for training neural network models with noisy multi-label data to solve the technical problem of training classification networks with strong decision-making capabilities in noisy multi-label learning scenarios.

[0010] The technical solution of this invention is as follows:

[0011] A method for training a neural network model on noisy multi-label data includes the following steps:

[0012] A "clean" sample set (hereinafter referred to as the clean sample set) is selected for each category using a sample selection algorithm as the meta dataset, and the category-dependent label noise transition matrix is ​​estimated.

[0013] The class-dependent label noise transition matrix is ​​used to initialize some parameters in the instance feature-dependent label noise transition matrix network;

[0014] Based on the statistical consistency of label noise learning loss, a meta-learning algorithm is used to simultaneously learn the network parameters of the label noise transition matrix, the data imbalance parameter, and the parameters of the multi-label classification neural network, which are dependent on the instance features.

[0015] Furthermore, the sample selection algorithm is an algorithm based on the memory effect of deep neural networks. This invention trains a classifier with a standard multi-label classification loss for several rounds on a noisy training sample set, and then executes the sample selection algorithm to obtain a selected clean sample set for each class label as the meta-dataset. Based on the obtained clean sample set and the existing noisy data, this invention utilizes the mismatch in label correlation to estimate the class-dependent label noise transition matrix.

[0016] Furthermore, by using the obtained category-dependent label noise transition matrix to initialize some parameters in the instance feature-dependent label noise transition matrix network, a good parameter initial point is provided for learning and optimizing the network.

[0017] Furthermore, this invention applies the statistical consistency algorithm in noisy single-label scenarios to each binary classification problem obtained from multi-label task decomposition, and uses the selected clean sample set as the meta-dataset. Through the meta-learning algorithm, it simultaneously learns the network parameters of the label noise transition matrix, the data imbalance parameter, and the multi-label classification neural network parameters that the instance features depend on, which can simultaneously alleviate the impact of label noise and data imbalance.

[0018] A neural network model training device for noisy multi-label data, comprising:

[0019] The sample selection and category-dependent label noise transition matrix estimation module is used to select a clean set of samples for each category as the meta-dataset through a sample selection algorithm, and to estimate the category-dependent label noise transition matrix.

[0020] The instance feature-dependent label noise transition matrix initialization module is used to initialize some parameters in the instance feature-dependent label noise transition matrix network using the category-dependent label noise transition matrix.

[0021] The meta-learning training module is used for learning the label noise loss based on statistical consistency. It employs a meta-learning algorithm to simultaneously learn the network parameters of the label noise transition matrix, the data imbalance parameters, and the parameters of the multi-label classification neural network, which are dependent on the instance features.

[0022] Compared with the prior art, the positive effects of the present invention are as follows:

[0023] 1) Use the clean sample set obtained by sampling samples from each category as the meta-dataset for noisy multi-label learning;

[0024] 2) The obtained category-dependent label noise transition matrix is ​​used to initialize some parameters in the instance feature-dependent label noise transition matrix network, providing a good parameter initial point for learning and optimizing the instance feature-dependent label noise transition matrix.

[0025] 3) Innovatively, meta-learning algorithms are used to learn and optimize the label noise transition matrix of instance features in a data-driven manner. By unifying the network parameters of the label noise transition matrix, which are dependent on instance features, the data imbalance parameters, and the parameters of the multi-label classification neural network under a unified framework, the impact of label noise and data imbalance can be mitigated simultaneously. Attached Figure Description

[0026] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0027] Figure 1 This is a flowchart of the method of the present invention.

[0028] Figure 2 This is a network structure diagram of the tag noise transfer matrix, which is a feature-dependent feature of the present invention.

[0029] Figure 3 This is a flowchart of the inner training process of the meta-learning algorithm of the present invention.

[0030] Figure 4 This is a flowchart of the outer training process of the meta-learning algorithm of the present invention.

[0031] Figure 5 A flowchart illustrating the application of this invention in the context of clothing attribute classification. Detailed Implementation

[0032] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0033] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0034] According to one aspect of the present invention, a method embodiment for training a model is provided. The method flow of the present invention is as follows: Figure 1 As shown, the method framework of this invention comprises sample selection and category-dependent label noise transition matrix estimation, instance feature-dependent label noise transition matrix initialization, and a meta-learning training module for noisy multi-label scenarios.

[0035] I. Sample selection and category-dependent label noise transition matrix estimation

[0036] Based on existing noisy multi-label data, in order to construct the metadata required for the meta-learning algorithm, this invention utilizes a sample selection algorithm to obtain data with the cleanest possible labels as the meta-dataset. Specifically, this invention uses noisy multi-label data D... t The neural network model f is preheated and trained for several epochs (e.g., 5-20 epochs), using the following standard multi-label binary classification learning loss function:

[0037]

[0038] Where f is a multi-label prediction neural network with a ResNet structure; f j (X) represents the prediction result of the j-th class obtained by the model through nonlinear mapping of the input data features X, where X represents the input feature vector; The noise label vector of the sample. q represents the noise label of class j; l represents the total number of classes; Y represents the binary cross-entropy loss; and Y represents the clean label vector of the sample.

[0039] After preheating training, the neural network model acquires a certain discriminative ability. This invention models the distribution of loss for each sample of each class j using a Gaussian mixture model (GMM), extracts a subset of samples with smaller losses, and obtains a selected clean sample set for each class label j.

[0040] Then, to facilitate subsequent learning of the label noise transition matrix network dependent on instance features, this invention utilizes the label correlation between multiple labels to estimate the category-dependent label noise transition matrix, i.e. Where Y j Let v be the clean label for class j, and v and k be 0 or 1, representing the negative or positive class. Specifically, some label correlations that shouldn't exist in practice are included in noisy multi-label learning. For example, in real multi-label data, "fish" and "water" always appear together, as do "bird" and "sky." However, due to labeling errors, there might be a slight correlation between "fish" and "sky," which is unrealistic. Therefore, the mismatch between metadata and label correlations in noisy data can be used for estimation.

[0041] The innovation of this step lies in: using a sample selection algorithm to obtain a set of selected clean samples for each label class as the meta-dataset for the meta-learning algorithm for noisy multi-label scenarios, and using it to estimate the label noise transition matrix of category dependence.

[0042] II. Initialization of the Label Noise Transition Matrix Dependent on Instance Features

[0043] To alleviate the difficulty of learning instance feature-dependent label noise transition matrix networks in noisy multi-label learning, this invention uses the estimated category-dependent label noise transition matrix to initialize some parameters of instance feature-dependent label noise transition matrix networks.

[0044] The structure of the instance feature-dependent label noise transfer matrix network is as follows: Figure 2 As shown, the network structure of the label noise transfer matrix network for category j consists of a backbone network g(.) and a linear layer L. j (.), the backbone network takes instance features x as input and outputs noise pattern representation g(x), while the linear layers take noise pattern representation g(x) as input and output. The output is the label noise transition matrix that depends on the instance features of class j. in Let f represent the set of predicted outputs of classifier f for features x in all classes except the j-th class.

[0045] Let the linear layer L j If the input variable for (.) is z, then it can be represented as L j (z) = az + b, where a and b are the learnable parameters of the linear layer. For ease of learning, during initialization, this invention initializes parameter a with a normal distribution of mean 0 and variance 0.01. For parameter b, this invention utilizes the estimated category-dependent label noise transition matrix of the j-th class, i.e. Initialize it. Such initialization ensures that the label noise transition matrix, which depends on the instance features, has an initial solution with relatively small loss, which is beneficial for subsequent optimization.

[0046] The innovation of this step lies in: using the category-dependent label noise transition matrix to initialize some parameters of the instance feature-dependent label noise transition matrix, thus alleviating the problem of the instance feature-dependent label noise transition matrix being difficult for the network to learn.

[0047] III. Meta-learning training framework for noisy multi-label scenarios

[0048] To alleviate the learning difficulties caused by the coupling between the data imbalance parameters and the label noise transition matrix parameters in noisy multi-label scenarios, this invention proposes a meta-learning training framework for noisy multi-label scenarios.

[0049] First, given input X, this invention decomposes the task into q conditionally independent binary classification problems, namely, predicting P(Y). j The terms |X) are independent of each other.

[0050] Secondly, this invention applies the statistically consistent label noise learning loss from the single-label case to each binary classification problem, summing the results to obtain the final loss for learning. Without loss of generality, a reweighting algorithm is used here as the statistically consistent single-label noise learning loss for learning the neural network classifier:

[0051]

[0052] Among them, L j Let be the loss function for the j-th class, which can be replaced by other statistically consistent binary classification noise label learning losses, where n is the total number of samples. Let be the noise label of the j-th class for the i-th sample;

[0053]

[0054] in, To transform the transition matrix T j The elements in (x) with the subscript ik represent the probability of flipping the j-th class label of sample x from value i to value k, where i and k = 0 or 1;

[0055] Among them, w j This is a data imbalance parameter used to mitigate data imbalance problems.

[0056] In this framework, it is necessary to simultaneously learn the noise transition matrix T that depends on the instance features. j (.), Data imbalance parameter w j and neural network classifier fj If the (.) parameter is directly minimized, there are multiple feasible solutions, which cannot guarantee the learning effect.

[0057] To address this problem, this invention transforms the learning problem into the following two-level optimization problem (e.g.) Figure 3 and Figure 4 And use meta-learning for training:

[0058]

[0059]

[0060] Among them, w * T is the learned optimal data imbalance parameter. * f is the learned optimal transition matrix. * Let w be the learnable optimal classifier, T be the learnable optimal transition matrix, and f be the learnable classifier. The dataset of class j selected previously using Gaussian mixture model A collection of The original noisy multi-label training set is used. In this two-layer optimization problem, the outer optimization layer uses a selected clean sample set as the meta-dataset to learn the noisy transition matrix T that depends on instance features. j (.) and the data imbalance parameter w j The inner optimization layer learns a neural network classifier f using a statistically consistent loss on the original noisy multi-label training set based on the learned noise transition matrix and data imbalance parameters. j The parameter of (.).

[0061] The innovation of this method module lies in: using a statistically consistent label noise learning loss, and employing a meta-learning algorithm to simultaneously learn the network parameters of the label noise transition matrix, the data imbalance parameters, and the parameters of the multi-label classification neural network that are dependent on the instance features.

[0062] By employing the above-mentioned technical solution of this invention, learning under inaccurate supervised information data is achieved, solving the following two problems. First, the problem of estimating the instance feature-dependent label noise transition matrix in noisy multi-label scenarios: Directly optimizing the instance feature-dependent label noise transition matrix without any assumptions is an ill-conditioned problem. To address this, this invention utilizes a clean sample set obtained through sample selection to provide additional information for learning. Furthermore, to further promote learning, this invention innovatively utilizes a class-dependent label noise transition matrix to initialize the instance feature-dependent label noise transition matrix. Second, the coupling problem between data imbalance parameters and label noise transition matrix parameters in noisy multi-label scenarios: Since statistical consistency algorithms require a good fit to the noise posterior probability during the learning process, class imbalance and positive / negative sample imbalance in multi-label scenarios often severely affect the learning of the noise posterior probability. If only the label noise transition matrix is ​​learned, the coupling effect of data imbalance will lead to inaccurate learning. To better alleviate this problem, this invention employs a meta-learning framework to simultaneously learn the network parameters of the instance feature-dependent label noise transition matrix, the data imbalance parameters, and the parameters of the multi-label classification neural network.

[0063] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0064] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention. For example:

[0065] One embodiment of the present invention provides a neural network model training apparatus for noisy multi-label data, comprising:

[0066] The sample selection and category-dependent label noise transition matrix estimation module is used to select a clean set of samples for each category as the meta-dataset through a sample selection algorithm, and to estimate the category-dependent label noise transition matrix.

[0067] The instance feature-dependent label noise transition matrix initialization module is used to initialize some parameters in the instance feature-dependent label noise transition matrix network using the category-dependent label noise transition matrix.

[0068] The meta-learning training module is used for learning the label noise loss based on statistical consistency. It employs a meta-learning algorithm to simultaneously learn the network parameters of the label noise transition matrix, the data imbalance parameters, and the parameters of the multi-label classification neural network, which are dependent on the instance features.

[0069] For the specific implementation process of each module, please refer to the description of the method of the present invention above.

[0070] Another embodiment of the present invention provides a computer device (mobile phone, computer, server, or network device, etc.) including a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the steps of the method of the present invention.

[0071] Another embodiment of the present invention provides a computer-readable storage medium (ROM / RAM, magnetic disk, optical disk, etc.) storing a computer program, which, when executed by a computer, implements the various steps of the method of the present invention.

[0072] The following is combined Figure 5 The application of the present invention in specific scenarios is described.

[0073] With the diverse forms and attributes of clothing today, effectively identifying multiple attributes has become a key focus for some e-commerce businesses. In recent years, training models using deep neural networks (DNNs) on clean, supervised data has yielded good results. However, this training method faces difficulties in many practical applications. On the one hand, the data labeling process requires extremely high manpower and time costs; on the other hand, many clothing attributes are highly complex and difficult to discern, and even manual labeling often suffers from significant inaccuracies. This invention relates to a neural network model training method for noisy, multi-labeled data, which can effectively solve this problem. Figure 5 This invention relates to a method for training a neural network model for noisy multi-label data. The method involves inputting low-dimensional features obtained from clothing image data through a feature extraction network and noisy multi-attribute labels into the neural network model. This method can train an accurate neural network model, thereby enabling accurate attribute classification and retrieval.

[0074] The parts of the technical solutions of the present invention that are not described in detail can be implemented using existing technologies.

[0075] In this invention, the sample selection algorithm, statistical consistency loss, neural network architecture, and meta-learning optimization method are not limited to the small-loss selection algorithm, reweighting algorithm, ResNet network, and SGD optimization method based on the Gaussian mixture model described in this invention. Instead, specific sample selection algorithms, statistical consistency losses, network architectures, and optimization methods can be designed and selected according to specific business scenarios. The cross-entropy loss function defined in this invention can be replaced with other general classification loss functions.

[0076] Obviously, the embodiments described above are merely some, not all, embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

Claims

1. An attribute classification and retrieval method based on neural network model training of noisy multi-label data, characterized by, The method comprises the following steps: The low-dimensional features obtained by the image data through the feature extraction network and the noisy multi-attribute label are taken as inputs, a neural network model training method for noisy multi-label data is used, a neural network model is trained, and the trained neural network model is used for accurate attribute classification and retrieval; The neural network model training method for noisy multi-label data comprises the following steps: A clean sample set is selected as a meta data set for each category by a sample selection algorithm, and a category-dependent label noise transfer matrix is estimated; Part of the parameters in the instance feature-dependent label noise transfer matrix network is initialized by using the category-dependent label noise transfer matrix; Based on the statistical consistency of the label noise learning loss, the meta learning algorithm is used to simultaneously learn the instance feature-dependent label noise transfer matrix network parameters, the data imbalance parameters and the multi-label classification neural network parameters; The following steps are used to obtain the clean sample set: In the noisy multi-label data D t The pre-heating training neural network model f for several rounds, which uses the following standard multi-label binary classification learning loss: wherein f is a multi-label prediction neural network, the structure of which is ResNet; f is used to j (X) represents the prediction result of the jth class obtained by the model performing nonlinear operation on the input X, wherein X represents an input feature vector; is a noise label vector of the sample, is the noise label of the jth class; q is the total number of classes; l is a binary cross-entropy loss; By using a Gaussian mixture model to model the distribution of the loss of each sample of each class j, a subset of samples with smaller loss is extracted, and a clean sample set of each class label j is obtained The learning problem is converted into the following double optimization problem by using the meta learning algorithm to simultaneously learn the instance feature-dependent label noise transfer matrix network parameters, the data imbalance parameters and the multi-label classification neural network parameters, and the meta learning is used for training: where w * is the learned optimal data imbalance parameter, T * is the learned optimal transition matrix, and f * is the learned optimal classifier. is the union of the selected j-class data sets using Gaussian mixture model, is the original noisy multi-label training set; in this two-level optimization problem, the outer optimization learns the instance feature dependent noise transition matrix T j (·) and the data imbalance parameter w j using the selected "clean" sample set as the meta data set, and the inner optimization learns the parameters of the neural network classifier f j (·) using the learned noise transition matrix and the data imbalance parameter on the original noisy multi-label training set with statistically consistent loss.

2. The method of claim 1, wherein, The clean sample set is selected as a meta data set for each category by a sample selection algorithm, and a category-dependent label noise transfer matrix is estimated.

3. The method of claim 1, wherein, The clean sample set is selected as a meta data set for each category by a sample selection algorithm, and a category-dependent label noise transfer matrix is estimated.

4. The method of claim 1, wherein, Part of the parameters in the instance feature-dependent label noise transfer matrix network is initialized by using the category-dependent label noise transfer matrix.

5. The method of claim 1, wherein, Part of the parameters in the instance feature-dependent label noise transfer matrix network is initialized by using the category-dependent label noise transfer matrix. Let L be the last linear layer of the transition matrix network j The input variable of (.) is z and is denoted by L j (z) = az + b, where a and b are the learnable parameters of this linear layer; The following steps are used to obtain the clean sample set:

6. An attribute classification and retrieval apparatus based on training of a neural network model using noisy multi-label data using the method according to any one of claims 1 to 5, characterized by, The learning problem is converted into the following double optimization problem by using the meta learning algorithm to simultaneously learn the instance feature-dependent label noise transfer matrix network parameters, the data imbalance parameters and the multi-label classification neural network parameters, and the meta learning is used for training: The clean sample set is selected as a meta data set for each category by a sample selection algorithm, and a category-dependent label noise transfer matrix is estimated. Part of the parameters in the instance feature-dependent label noise transfer matrix network is initialized by using the category-dependent label noise transfer matrix. Part of the parameters in the instance feature-dependent label noise transfer matrix network is initialized by using the category-dependent label noise transfer matrix. The following steps are used to obtain the clean sample set: The learning problem is converted into the following double optimization problem by using the meta learning algorithm to simultaneously learn the instance feature-dependent label noise transfer matrix network parameters, the data imbalance parameters and the multi-label classification neural network parameters, and the meta learning is used for training: The clean sample set is selected as a meta data set for each category by a sample selection algorithm, and a category-dependent label noise transfer matrix is estimated. The device takes low-dimensional features obtained by a feature extraction network and a noisy multi-attribute label as input, uses a sample selection and class-dependent label noise transfer matrix estimation module, an instance feature-dependent label noise transfer matrix initialization module and a meta-learning training module to train a neural network model, and uses the trained neural network model for accurate attribute classification and retrieval.

7. A computer device, comprising: The device comprises a memory and a processor, the memory stores a computer program configured to be executed by the processor, and the computer program comprises instructions for executing the method of any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by a computer to implement the method of any one of claims 1-5.

Images