Image classification method and system based on similarity uncertainty optimization

Abstract

The invention discloses an image classification method and system based on similarity uncertainty optimization, and belongs to the technical field of computer vision. Comprises: acquiring a to-be-processed image; processing the to-be-processed image through a trained image classification model to obtain an image classification result; and when the image classification model is trained, quantifying the similarity uncertainty between a difficult sample in a training set containing noise and a dynamic class prototype, and performing back propagation optimization on a constructed network in combination with classification loss and divergence evaluation loss calculated through a clean sample in the training set. According to the method, more discriminative information can be learned from samples with relatively high uncertainty, the robustness of the model is improved, and the problem that noise data affects model prediction stability during existing model training is solved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to an image classification method and system based on similarity uncertainty optimization. Background Art

[0002] The statements in this section merely mention background art related to the present invention and do not necessarily constitute prior art.

[0003] Noisy label learning is an important research direction in machine learning, especially in supervised learning tasks. Traditional supervised learning relies on high-quality labeled datasets, but in practical applications, it is often very difficult to obtain completely accurate labeled data, especially for large-scale datasets.

[0004] Due to human error, sensor failure, or limitations of the automatic annotation system, data labels may contain noise, that is, labels that do not match the true category. This noise not only affects the generalization ability of the model, but may also cause the model to overfit to the wrong labels, thereby reducing its performance.

[0005] The development of noisy label learning can be divided into several stages. Early research focused on reducing the impact of noise on model training by modeling it, usually assuming that noise is generated from a certain distribution, and using statistical methods or Bayesian methods to construct a noise transfer matrix to infer the true label. Subsequently, with the rise of deep learning, noisy label learning methods gradually turned to more complex models and techniques, such as loss function modification, data cleaning, and model robustness enhancement. Some methods reduce the impact of noise on gradients by adjusting the loss function, while another type of method improves data quality by cleaning data and removing or correcting incorrect labels.

[0006] In recent years, dual-model training, semi-supervised learning, and self-supervised learning techniques have also made significant progress in noisy label learning. These methods use unlabeled data or mutual learning between models to alleviate the impact of noise. However, none of the above methods consider the impact of inherent noise that may exist in the data, which often leads to overfitting or performance degradation of the model, thereby affecting the accuracy and robustness of the model.

[0007] In addition, existing methods attempt to address the above issues by selecting or re-weighting samples, but they ignore the impact of perceptual ambiguity and bias caused by the inherent noise contained in the dataset and rely heavily on the softmax classifier, ignoring the exploration of knowledge in low-quality samples. Summary of the invention

[0008] In order to address the deficiencies of the prior art, the present invention provides an image classification method, system, electronic device, computer-readable storage medium and computer program product based on similarity uncertainty optimization, which utilizes the inherent noise in uncertain quantization data sets to improve the accuracy and robustness of image classification.

[0009] In a first aspect, the present invention provides an image classification method based on similarity uncertainty optimization;

[0010] An image classification method based on similarity uncertainty optimization, comprising:

[0011] Get the image to be processed;

[0012] Processing the image to be processed by using a trained image classification model to obtain an image classification result;

[0013] When training the image classification model, the uncertainty of the similarity between the difficult samples in the noisy training set and the dynamic class prototype is quantified, and the constructed network is back-propagated and optimized in combination with the classification loss and the divergence evaluation loss calculated by the clean samples in the training set.

[0014] In some embodiments, the quantifying the similarity uncertainty between the difficult samples in the noisy training set and the dynamic class prototype specifically includes:

[0015] Based on the difficult samples and the dynamic class prototype, a similarity uncertainty distribution is constructed by combining cosine similarity calculation and graph convolutional neural network;

[0016] Monte Carlo algorithm is used to sample from similarity uncertainty distribution and calculate similarity uncertainty loss;

[0017] The dynamic class prototype is constructed using clean samples.

[0018] In some embodiments, the divergence assessment loss when training the image classification model is calculated using clean samples in the training set as follows: the clean samples are processed through an MLP network to obtain a clean score corresponding to each of the clean samples, and the divergence assessment loss is calculated using dual network prediction divergence for supervision.

[0019] In some implementations, calculating the classification loss when training the image classification model is specifically as follows: performing dynamic label masking according to the confidence of each sample and calculating the classification loss;

[0020] The label corresponding to the difficult sample is a pseudo label corrected by linear addition of cosine similarity and dual network prediction.

[0021] In some embodiments, training the image classification model also includes: extracting features from the training set, inputting the extracted features into the GMM algorithm for processing, obtaining the confidence level corresponding to each sample, and dividing the samples based on the confidence level.

[0022] In some embodiments, back propagation optimization of the constructed network is specifically performed by linearly adding the classification loss, similarity uncertainty loss, and divergence evaluation loss, and calculating the gradient to perform back propagation optimization of the network.

[0023] In a second aspect, the present invention provides an image classification system based on similarity uncertainty optimization;

[0024] An image classification system based on similarity uncertainty optimization, comprising:

[0025] The acquisition module is configured to: acquire the image to be processed;

[0026] The image classification module is configured to: process the image to be processed by using a trained image classification model to obtain an image classification result;

[0027] When training the image classification model, the uncertainty of the similarity between the difficult samples in the noisy training set and the dynamic class prototype is quantified, and the constructed network is back-propagated and optimized in combination with the classification loss and the divergence evaluation loss calculated by the clean samples in the training set.

[0028] In a third aspect, the present invention provides an electronic device;

[0029] An electronic device comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-mentioned image classification method based on similarity uncertainty optimization.

[0030] In a fourth aspect, the present invention provides a computer-readable storage medium;

[0031] A computer-readable storage medium stores a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned image classification method based on similarity uncertainty optimization.

[0032] In a fifth aspect, the present invention provides a computer program product;

[0033] A computer program product comprises a computer program / instruction, which implements the steps of the above-mentioned image classification method based on similarity uncertainty optimization when executed by a processor.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. The technical solution provided by the present invention quantifies the similarity uncertainty of the fixed noise in the data set, and quantifies the uncertainty in the sample by modeling the similarity distribution between the difficult sample and the class prototype through cosine similarity and graph convolutional neural network, thereby helping the model to better utilize the information in low-quality samples.

[0036] 2. The technical solution provided by the present invention, in order to solve the problem of overconfidence of the network, implements a dynamic masking strategy for the prediction of each sample, and calculates the classification loss for the masked prediction label. The comprehensive loss optimizes the network and can improve the classification performance of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0038] Figure 1 A schematic diagram of a process flow provided by an embodiment of the present invention;

[0039] Figure 2 A detailed flowchart provided for an embodiment of the present invention;

[0040] Figure 3 A schematic diagram of a process for calculating similarity uncertainty loss provided by an embodiment of the present invention;

[0041] Figure 4 A schematic diagram of a prediction process of a divergence assessment loss provided by an embodiment of the present invention;

[0042] Figure 5 An example diagram of image classification provided by an embodiment of the present invention;

[0043] Figure 6 A schematic diagram of the process of performing sample division using the GMM algorithm provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0045] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "include" and "have" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0046] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0047] Terminology explanation:

[0048] Noisy label learning: Noisy label learning refers to the technology of dealing with label errors or inaccurate annotations in data sets in machine learning. Since real-world data sets may contain noisy labels, which will affect the performance of the model, noisy label learning reduces the negative impact of noisy labels on model training and improves the generalization ability and robustness of the model through methods such as algorithm adjustment, data cleaning, and model weighting.

[0049] Uncertainty: There are usually two types of uncertainty: one is structural uncertainty, which refers to problems such as poor model performance caused by the selected model structure and parameters. This uncertainty can often be solved by increasing the sample size. The other is cognitive uncertainty, which is caused by problems such as noise inherent in the data set.

[0050] Graph Convolutional Network: A model that specifically processes graph-structured data. It updates the representation of nodes by aggregating node features with information from neighboring nodes. It can effectively capture the relationship between nodes and edges in the graph and is widely used in fields such as social network analysis, recommendation systems, and bioinformatics.

[0051] ResNet18: A deep convolutional neural network based on residual network with 18 layers of depth. It alleviates the gradient vanishing problem in deep networks by introducing skip connections, making training deeper networks more efficient.

[0052] Class prototype: refers to a central point or feature vector representing a class or cluster, which is used to describe the typical characteristics or patterns of the class. Common forms of class prototypes include centroids or other features that can characterize the distribution of data within the cluster, helping to measure the similarity between new data points and the class.

[0053] Similarity: An indicator used to measure the similarity between two objects, usually expressed in numerical form. It is used to compare the features of objects. The higher the value, the more similar the objects are. Common calculation methods include Euclidean distance, cosine similarity, etc.

[0054] GMM: is a probabilistic model that models data as a mixture of multiple Gaussian distributions, commonly used in soft clustering. Through the expectation maximization (EM) algorithm, GMM can estimate the parameters of each Gaussian distribution and is widely used in tasks such as clustering and anomaly detection.

[0055] Clean samples: samples that are considered to be correctly labeled.

[0056] Difficult samples: Samples that are difficult to classify and whose labels are often wrong.

[0057] MLP (Multilayer Perceptron): It is a feedforward neural network consisting of an input layer, one or more hidden layers, and an output layer. The neurons in each layer are connected by weights and nonlinear activation functions. It is widely used in classification and regression tasks.

[0058] Embodiment 1

[0059] The existing image classification model suffers from overfitting due to the uncertainty of noise labels, which affects the accuracy and robustness of image classification. Therefore, the present invention provides an image classification method based on similarity uncertainty optimization.

[0060] Next, combine Figure 1-Figure 6 , a method for image classification based on similarity uncertainty optimization disclosed in this embodiment is described in detail. The method for image classification based on similarity uncertainty optimization includes the following steps:

[0061] S1. Obtain an image to be processed.

[0062] In this embodiment, the images to be processed are RGB images with a size of 32*32, and each image contains an object to be classified.

[0063] S2. Process the image to be processed through the trained image classification model to obtain the image classification result.

[0064] In this embodiment, the image classification model is a ResNet18 model, the last layer of which is a fully connected layer capable of classifying images; the input of the ResNet18 model is the image to be processed, and the output is the image classification result.

[0065] The image classification result means that the model outputs a prediction score for each image corresponding to each category. The prediction score indicates the probability that the image belongs to the corresponding category. The category with the largest probability is selected as the prediction result of the model, that is, the image is considered to belong to that category. The presence of label noise increases the uncertainty of the data, making it difficult for the model to distinguish which samples have accurate labels and which are wrong. Generally, noisy labels make the model's predictions for certain samples more unstable because the model may mistakenly learn the noisy labels as valid signals. This uncertainty comes from many aspects, including the noise in the data itself, the inaccurate labeling process, and the model's tendency to overfit on noisy samples.

[0066] This makes uncertainty quantification and modeling the key to solving the noisy label problem. Therefore, in this embodiment, during the model training process, this part of uncertainty is quantified and used to solve the noisy label problem.

[0067] Therefore, in this embodiment, two initial image classification models are constructed, and the architecture and parameters of the initial image classification models are the same. Then, the two initial image classification models are trained in parallel based on the noise label learning method with similarity uncertainty optimization, and the training methods are the same. Finally, the trained models are evaluated, and the initial image classification model with higher classification accuracy is used as the final image classification model.

[0068] As an implementation method, the specific process of training the image classification model is as follows:

[0069] Step 1: Build two parallel neural networks.

[0070] Step 2: Obtain a noisy data set as a training set, and extract features from the training set using the feature extractor in the initial image classification model.

[0071] In this embodiment, the feature extractor is a ResNet18 network.

[0072] Step 3: Input the extracted features into the GMM algorithm for processing. The GMM algorithm calculates a confidence for each sample. When the confidence is higher than the given threshold, these samples are classified as clean samples, and the remaining samples are classified as difficult samples. For the clean sample set, steps 4 and 5 are executed in parallel; for the difficult sample set, steps 6 and 7 are executed in parallel.

[0073] Combination Figure 6 Specifically, the extracted features are sample normalized to obtain the initialized mean vector, covariance matrix and mixing coefficient, and then it is determined whether the preset number of iterations has been reached; if so, the confidence score of each sample is output; if not, the EM algorithm is used for iterative calculation.

[0074] In this embodiment, no improvement is made to the existing GMM algorithm, and the process is automatically completed by calling the existing algorithm, which will not be described in detail here.

[0075] Step 4: Input the clean sample set into the MLP network, and calculate a clean score for each clean sample through the MLP network as the probability that the sample may be a clean sample. Use the prediction of the dual network to calculate the JS divergence as the probability that the sample is approximately clean, and use this probability as supervision to calculate the prediction divergence evaluation loss.

[0076] In this embodiment, in order to evaluate the importance of different clean samples, an MLP network is used to process the clean sample set. However, due to the inherent noise in the data, the true probability of the sample is unknown, which makes MLP supervision difficult. To solve this problem, in this embodiment, it is proposed to approximate the true noise type of each clean sample through Jensen-Shannon (JS) divergence, and train MLP accordingly.

[0077] The training process has two networks trained in parallel. By using JS divergence to estimate the prediction inconsistency between the two networks as an approximation of the true probability of clean samples, the prediction divergence can be optimized more smoothly. The JS divergence calculation formula is expressed as follows:

[0078]

[0079] Where P(i) and Q(i) represent the prediction results of two parallel networks respectively.

[0080] Step 5: Perform a certain degree of dynamic label masking on the prediction of each sample in the clean sample set according to the confidence level and calculate the classification loss. The classification loss function of the clean sample set is as follows:

[0081]

[0082] Among them, L ce is the cross entropy loss commonly used in the field of noisy labels, Indicates the label, Represents the model's predictions.

[0083] Models generally have the problem of overconfident predictions, that is, the model shows an unreasonably high confidence in its wrong prediction results when predicting. Specifically, the model may give very high prediction probabilities on certain input samples, even if these predictions are wrong. Therefore, in this embodiment, the model's predictions are masked to a certain extent according to the confidence level to combat the model's overconfidence behavior.

[0084] Step 6: Calculate the class prototype based on the clean sample, quantify the similarity uncertainty between the difficult sample and the dynamic class prototype, and calculate the similarity uncertainty loss.

[0085] In label noise learning (LNL), noisy labels usually come from low-quality samples. Existing methods often ignore the differences between these samples and class prototypes, which affects the overall performance. In this embodiment, samples in difficult datasets are regarded as low-quality data, and the uncertainty of the similarity between difficult samples and class prototype pairs is characterized to suppress perceptual ambiguity and bias. Specifically, the similarity representation between difficult samples and prototype pairs is converted from a deterministic scalar to a probabilistic representation, so that this uncertainty is naturally incorporated into the loss calculation for better model optimization.

[0086] For example, the specific process is as follows:

[0087] (1) The class prototype is calculated based on the clean sample, and the cosine similarity between the difficult sample and the class prototype is calculated as the mean of their similarities.

[0088] Specifically, the class prototype calculation based on the clean sample is as follows: the confidence obtained in step 3 is used as the weight coefficient, and the class prototype calculated in the previous training stage is dynamically weighted to obtain the class prototype. It can be expressed as:

[0089]

[0090] In the formula, is the class prototype of the kth category in the bth training stage, is the number of elements belonging to the kth class in the clean set, w i is the confidence of the corresponding sample, f φ (x i ) is the characteristic of the corresponding sample, is the class prototype of the kth category in the b-1th training stage.

[0091] (2) Use a graph convolutional neural network to calculate the variance between difficult samples and each class, and use the mean and variance to construct the similarity uncertainty distribution.

[0092] Specifically, the mean of the similarity between the difficult sample and the class prototype in step (1) is input as a node into the graph convolutional neural network, processed in sequence by multiple convolutions, batch normalization and activation functions, and then passed through a softplus layer to ensure that the output is a positive value. The output result is normalized so that the output result of GCN is the variance of the difficult sample corresponding to each category.

[0093] A normal distribution is constructed based on the mean and variance of each type corresponding to the difficult samples obtained above, and the normal distribution is used as the similarity uncertainty distribution.

[0094] The normal distribution is a distribution that satisfies N(μ,σ 2), where μ is the mean, σ 2 is the variance.

[0095] (3) Use the Monte Carlo algorithm to sample from the similarity uncertainty distribution and calculate the similarity uncertainty loss.

[0096] In this embodiment, the mean and variance of each type corresponding to the obtained difficult samples are used as the similarity and uncertainty of the samples respectively. However, the gradient calculated directly from the distribution constructed by the mean and variance is 0, which is difficult to be used for optimizing the network.

[0097] Therefore, in this step, the mean and variance of each type corresponding to the obtained difficult samples are reparameterized to obtain a quantized uncertainty value, that is, a differentiable normal distribution; then, the differentiable normal distribution is sampled by the Monte Carlo algorithm to calculate the uncertainty loss; the above process is repeated multiple times, and the average of all the uncertainty losses obtained is taken as the final uncertainty loss.

[0098] The uncertainty value is expressed as:

[0099] s=μ+λ*σ 2 ;

[0100] Among them, s represents the uncertainty value, μ represents the calculated mean, and σ 2 represents the calculated variance, and λ represents a random variable from a standard normal distribution.

[0101] The calculation formula of uncertainty loss is as follows:

[0102]

[0103] In the formula, x represents the sample, k is the category to which the sample belongs, and s k,t and j,t It represents the uncertainty value obtained by sampling through the Monte Carlo algorithm, and N is the total number of categories.

[0104] Step 7: Use the linear sum of cosine similarity and dual model predictions as the pseudo-label of the difficult sample, perform dynamic label masking and calculate the classification loss.

[0105] Among them, the classification loss function of difficult samples is expressed as follows:

[0106]

[0107] In the formula, represents the obtained pseudo label, Represents the model's predictions.

[0108] Based on this, cosine similarity is used to correct pseudo labels for difficult samples to obtain more realistic pseudo labels.

[0109] Step 8: Linearly sum all losses and calculate the gradient to perform back-propagation optimization on the network.

[0110] Step 9: Determine whether the set number of cycles has been reached. If it has been completed, the network at this time is the trained initial image classification model; otherwise, continue the optimization cycle.

[0111] Step 10: The one with higher accuracy among the two trained initial image classification models is used as the final image classification model.

[0112] In summary, in this embodiment, first, the GMM method is used to divide the samples into clean samples and hard samples; then, the dynamic class prototype is constructed using the clean samples, and the uncertainty is quantified based on the similarity between the class prototype and the clean samples, thereby suppressing the perceptual ambiguity and bias in the noise data. Among them, the importance of clean samples is evaluated using the prediction divergence evaluation, and stricter constraints are imposed on the model. The model is optimized using the uncertainty perception loss function, the masked classification loss function, and the prediction divergence evaluation loss function to make the model more robust to noise data.

[0113] Embodiment 2

[0114] This embodiment discloses an image classification system based on similarity uncertainty optimization, including:

[0115] The acquisition module is configured to: acquire the image to be processed;

[0116] The image classification module is configured to: process the image to be processed by using a trained image classification model to obtain an image classification result;

[0117] When training the image classification model, the uncertainty of the similarity between the difficult samples in the noisy training set and the dynamic class prototype is quantified, and the constructed network is back-propagated and optimized in combination with the classification loss and the divergence evaluation loss calculated by the clean samples in the training set.

[0118] It should be noted that the acquisition module and the image classification module correspond to the steps in Embodiment 1, and the examples and application scenarios implemented by the modules and the corresponding steps are the same, but are not limited to the contents disclosed in Embodiment 1. It should be noted that the modules as part of the system can be executed in a computer system such as a set of computer executable instructions.

[0119] Embodiment 3

[0120] Embodiment 3 of the present invention provides an electronic device, including a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps of the above-mentioned image classification method based on similarity uncertainty optimization are completed.

[0121] Embodiment 4

[0122] Embodiment 4 of the present invention provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps of the above-mentioned image classification method based on similarity uncertainty optimization are completed.

[0123] Embodiment 5

[0124] Embodiment 5 of the present invention provides a computer program product, including a computer program / instruction, which implements the steps of the above-mentioned image classification method based on similarity uncertainty optimization when executed by a processor.

[0125] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0126] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0127] These computer program instructions can also be loaded onto a computer or other programmable data processing device, and a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0128] The description of each embodiment in the above embodiments has different emphases. For parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0129] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An image classification method based on similarity uncertainty optimization, characterized in that: include: Get the image to be processed; Processing the image to be processed by using a trained image classification model to obtain an image classification result; When training the image classification model, the uncertainty of the similarity between the difficult samples in the noisy training set and the dynamic class prototype is quantified, and the constructed network is back-propagated and optimized in combination with the classification loss and the divergence evaluation loss calculated by the clean samples in the training set.

2. The image classification method based on similarity uncertainty optimization according to claim 1, characterized in that: The uncertainty of the similarity between the difficult samples in the noisy training set and the dynamic class prototype specifically includes: Based on the difficult samples and the dynamic class prototype, a similarity uncertainty distribution is constructed by combining cosine similarity calculation and graph convolutional neural network; Monte Carlo algorithm is used to sample from similarity uncertainty distribution and calculate similarity uncertainty loss; The dynamic class prototype is constructed using clean samples.

3. The image classification method based on similarity uncertainty optimization according to claim 1, characterized in that: The divergence evaluation loss when training the image classification model is calculated by using clean samples in the training set. Specifically, the clean samples are processed by the MLP network to obtain the clean score corresponding to each clean sample, and the divergence evaluation loss is calculated by using the dual network prediction divergence as supervision.

4. The image classification method based on similarity uncertainty optimization according to claim 1, characterized in that: The classification loss when the image classification model is trained is calculated as follows: dynamic label masking is performed according to the confidence of each sample and the classification loss is calculated; The label corresponding to the difficult sample is a pseudo label corrected by linear addition of cosine similarity and dual network prediction.

5. The image classification method based on similarity uncertainty optimization according to claim 1, characterized in that: Training the image classification model also includes: extracting features from the training set, inputting the extracted features into the GMM algorithm for processing, obtaining the confidence level corresponding to each sample and dividing the samples based on the confidence level.

6. The image classification method based on similarity uncertainty optimization according to claim 1, characterized in that: The back propagation optimization of the constructed network is specifically as follows: linearly add the classification loss, similarity uncertainty loss and divergence evaluation loss, and calculate the gradient to back propagate and optimize the network.

7. An image classification system based on similarity uncertainty optimization, characterized in that: include: The acquisition module is configured to: acquire the image to be processed; The image classification module is configured to: process the image to be processed by using a trained image classification model to obtain an image classification result; When training the image classification model, the uncertainty of the similarity between the difficult samples in the noisy training set and the dynamic class prototype is quantified, and the constructed network is back-propagated and optimized in combination with the classification loss and the divergence evaluation loss calculated by the clean samples in the training set.

8. An electronic device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the image classification method based on similarity uncertainty optimization according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by a processor, the steps of the image classification method based on similarity uncertainty optimization described in any one of claims 1 to 6 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the image classification method based on similarity uncertainty optimization described in any one of claims 1 to 6 are implemented.

Images