Semi-supervised image classification method, system, computer device and medium

Abstract

The application provides a semi-supervised image classification method, system, computer device and medium, and belongs to the technical field of image classification. The method performs image classification by constructing and evolving a double-track category prototype system (including anchor prototypes and evolved prototypes), uses labeled images to stably update the anchor prototypes, and screens high-quality unlabeled images to update the evolved prototypes through a multi-stage sample quality filter. The classifier is based on the dynamic fusion and sparsification construction of the anchor prototypes and the evolved prototypes, and optimizes the model through a composite objective function of supervised loss, unsupervised loss and proxy loss. Through the double-track prototype mechanism and sample filtering, the application effectively suppresses the semantic drift caused by pseudo-label noise, and the design of sparse prototype matching changes the classification process into transparent geometric measurement, improves the robustness and explainability of the model, and provides strong technical support for applications in medical, financial and other high-reliability fields.

Description

Technical Field

[0001] This invention relates to the field of image classification technology, and in particular to semi-supervised image classification methods, systems, computer devices, and media. Background Technology

[0002] Deep learning technology has achieved remarkable success in fields such as image recognition, but its widespread application still faces two major challenges: A strong dependence on large-scale labeled data. The superior performance of deep models relies on massive amounts of high-quality labeled data. However, in specialized fields such as medical imaging and industrial quality inspection, data labeling is extremely costly, leading to a scarcity of labeled samples. In such low-label scenarios, the performance of fully supervised models degrades sharply. While semi-supervised learning aims to alleviate this problem by utilizing unlabeled data, existing methods still face significant bottlenecks in practical applications.

[0003] The black-box nature of model decision-making processes. Traditional deep learning models typically make decisions through complex linear classification layers, resulting in a lack of transparency and interpretability in their reasoning process. This lack of interpretability poses serious security risks in high-risk areas (such as financial risk control and autonomous driving), and also makes it difficult to diagnose and repair models when they encounter adversarial attacks.

[0004] Specifically in the field of semi-supervised learning, current mainstream methods (such as FixMatch, UDA, etc.) generally rely on pseudo-labels and consistency regularization techniques, but still have the following limitations: Confirmation bias and error accumulation: The erroneous predictions generated by pseudo-labeling techniques in the early stages of training are continuously reinforced by subsequent iterations, leading to error solidification and semantic drift, which is particularly serious in difficult tasks such as fine-grained classification.

[0005] Insufficient robustness to noise and outlier samples: Existing methods lack effective sample filtering mechanisms, which can easily lead to the incorrect assignment of high-confidence pseudo-labels to noise or unseen categories (open set samples) in the data, thereby contaminating the entire learning process.

[0006] The feature space lacks structure: End-to-end implicit learning makes the feature space learned by the model not compact enough within classes and not separated enough between classes, making it difficult to cope with complex scenarios where intra-class differences are greater than inter-class differences, thus limiting the model's generalization ability in real-world environments.

[0007] Therefore, developing a new semi-supervised learning paradigm that can efficiently utilize unlabeled data while possessing strong robustness and decision interpretability has become a key issue that urgently needs to be addressed in this field. Summary of the Invention

[0008] The purpose of this invention is to provide a semi-supervised image classification method, system, computer device, and medium to solve problems such as performance degradation of image classification in scenarios with few labels, insufficient robustness to noise and abnormal samples, and lack of interpretability of the model decision process, thereby improving the classification accuracy and reliability of the model in real and complex environments.

[0009] To achieve the above objectives, the present invention provides a semi-supervised image classification method, comprising the following steps: Acquire image data, including labeled and unlabeled images; The image data is input into the image classification model to obtain the final classification result; The image classification model classifies images by constructing and evolving a dual-track category prototype system, including: The feature extraction module is used to extract feature vectors from the input image; The dual-track prototype evolution module is used to manage and update the anchor prototype and the evolving prototype; The classification and loss calculation module is used to fuse and sparsify the anchor prototype and the evolution prototype to generate a sparse fused prototype as a classifier, and calculate the loss based on the classifier to optimize the model.

[0010] Preferably, the dual-track prototype evolution module updates the anchor prototype and the evolving prototype in the following ways: The anchor prototype is updated using only the feature vectors of labeled image data via an exponential moving average. The evolutionary prototype is updated using feature vectors from high-quality unlabeled image data filtered through a multi-stage sample quality filter.

[0011] Preferred multi-stage sample quality filters include: High confidence screening steps: Screen unlabeled samples whose predicted probability is higher than a preset threshold; Category consistency check steps: Calculate the feature center of the unlabeled sample after screening within its pseudo-label category, and remove outlier samples that are too far from the center; Outlier samples are those whose features are far from the center of their pseudo-label category (mean distance within the category + twice the standard deviation of the distance within the category).

[0012] Preferably, the classification process of the classification and loss calculation module includes: The feature vector is input into the activation function to obtain the activated feature vector; The anchored prototype and the evolving prototype are dynamically weighted and fused into a dense fused prototype; The classification prediction result is obtained by calculating the dot product between the activated feature vector and the sparse fusion prototype.

[0013] Preferably, the loss calculation in the classification and loss calculation module includes: Supervised loss: calculated based on the prediction results and the ground truth labels of labeled images; Unsupervised loss: Pseudo-labels are generated based on predictions of weakly enhanced unlabeled images, and consistency loss is calculated between the prediction results of strongly enhanced unlabeled images; Proxy loss: calculated based on the mean square error between the feature vector and the corresponding target prototype in the sparse fusion prototype.

[0014] Preferably, the objects for calculating agency loss include: Feature vectors of labeled images; Feature vectors of high-quality unlabeled image data after being filtered through a multi-stage sample quality filter.

[0015] Preferably, the optimization process of the image classification model includes: The total loss is obtained by weighted summing of the supervised loss, the unsupervised loss, and the agency loss. The feature extraction module is optimized by backpropagation based on the total loss.

[0016] The present invention also provides a semi-supervised image classification system, comprising: The data input module is used to acquire labeled and unlabeled images; The feature extraction unit, used to extract feature vectors from an image, includes the following units: A dual-track prototype evolution unit is used to maintain and update the anchor prototype and the evolution prototype; The classification and loss calculation unit is used to construct the sparse fusion prototype and perform classification and loss calculation. The model optimization unit is used to optimize the model parameters based on the total loss.

[0017] The present invention also provides a computer device including a memory and a processor, the memory being used to store instructions and the processor being used to execute the instructions to implement the semi-supervised image classification method as described above.

[0018] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the semi-supervised image classification method as described above.

[0019] Therefore, the present invention employs the aforementioned semi-supervised image classification method, system, computer equipment, and medium, and the beneficial technical effects are as follows: (1) By establishing an anchor prototype that is stably updated by only a small amount of labeled data, a reliable learning benchmark is provided for the model. This design can effectively block the vicious cycle of "confirmation bias" in which false label errors in the early training are amplified during the iteration process, thereby significantly suppressing semantic drift and ensuring that the model performance remains stable even when labeled data is extremely scarce.

[0020] (2) Through an innovative multi-dimensional sample quality filter (integrating multiple criteria such as high confidence screening and class consistency check), high-quality unlabeled samples can be accurately identified and screened for training. This mechanism greatly reduces the interference of noisy data and open set samples on model training, avoids pseudo-label pollution, and thus maintains high performance even in real, noisy data environments.

[0021] (3) This invention abandons the traditional “black box” linear classification head and transforms the classification decision into a transparent, geometrically metric-based process (i.e., the similarity comparison between the feature vector and the sparse prototype). This makes the model’s decision basis no longer the difficult-to-interpret tens of thousands of dimensional parameters, but rather the intuitively understandable key features of each category that are most discriminative (corresponding to the non-zero dimension in the sparse prototype), providing a clear path for decision attribution and analysis, and meeting the needs of high-reliability application fields.

[0022] (4) By introducing a "surrogate loss" term, the feature vectors are explicitly forced to align with their corresponding class prototypes in space, thereby actively constructing a structured feature space with good intra-class compactness and inter-class separability. This not only directly improves classification accuracy, but more importantly, enhances the model's ability to generalize to unseen data. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall framework structure of the semi-supervised image classification method in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the workflow of a multi-stage sample quality filter used to screen high-quality unlabeled samples in an embodiment of the present invention. Figure 3 This is a flowchart of the overall framework in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the use of proxy loss for feature space alignment in an embodiment of the present invention; Figure 5 This is a flowchart of prototype-based classification and loss calculation in an embodiment of the present invention; Figure 6 This is a line graph showing the Top-1 accuracy performance of the method in this embodiment of the invention on the CIFAR-10 dataset under different numbers of labeled samples. Detailed Implementation

[0024] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0026] Example 1 This embodiment provides a semi-supervised image classification method, the complete process of which in one training iteration is as follows: Figure 1 and Figure 3 As shown. The core idea of ​​this method is to build a dynamic, robust, and interpretable classifier by evolving a set of dual-track prototype systems in parallel, and to optimize the model using a composite objective function consisting of supervised loss, unsupervised loss, and surrogate loss.

[0027] Step 1: Data input and feature extraction.

[0028] like Figure 1 As shown in the upper left part, at the beginning of a training iteration, the system first starts from the labeled dataset. and unlabeled datasets The data is sampled from the middle to form a data batch, in which, This represents a single labeled sample. Indicates a label, Indicates the first Each piece of labeled data This represents the total number of labeled samples. This represents a single unlabeled sample. Indicates the first Unlabeled data, This represents the total number of unlabeled data samples.

[0029] For unlabeled data, two different strengths of data augmentation will be performed: one is weak augmentation. (e.g., random flipping and cropping) to obtain unlabeled samples after weak enhancement. ,in This refers to unlabeled samples. Another type is strong enhancement. (e.g., RandAugment, Cutout, etc.) to obtain unlabeled samples after strong enhancement. .

[0030] Subsequently, ( , , The input is fed into a shared deep convolutional neural network as a feature extraction module. This network can be a commonly used backbone network in the field, such as the ResNet series (ResNet-34). The network performs forward propagation on the input image and outputs its representation in a high-dimensional feature space. To eliminate the influence of the modulus length and make subsequent similarity calculations more stable, the original feature vector output by the network is passed through an L2 normalization layer to obtain the feature vector z. The feature vector z is a D-dimensional real vector. ∈ This is the basis for all subsequent operations, and its calculation method is as follows: ; in, It is the original feature vector extracted from the input x.

[0031] Step 2: Dual-track prototype evolution.

[0032] like Figure 1 As shown in the upper part, this process runs in parallel with the main classification process, and its goal is to build and maintain a high-quality classifier. This module contains two core components: the anchor prototype and the evolution prototype.

[0033] Anchoring Prototype ( Update: Feature vectors of labeled samples It is directly used to update the anchor prototype. Specifically, for each category c appearing in the batch, the average of the features of all labeled samples belonging to that category is first calculated, and then this average is smoothly incorporated into the anchor prototype using an exponential moving average (EMA). In the corresponding category prototype, its update formula can be expressed as: ; in, Indicates the updated category The anchored prototype, Indicates the category before the update. The anchored prototype, This indicates that there are label data belonging to the category in the current batch. The sample index set, The momentum coefficient of the EMA is set to 0.99. This represents the feature vector of the k-th sample in the batch containing labeled samples.

[0034] Because the prototype is only affected by data with real labels, it provides an extremely stable and reliable benchmark ("anchor") for the entire model, thus effectively preventing semantic drift caused by the accumulation of pseudo-label noise in scenarios with few labels.

[0035] Evolutionary prototype ( Update: Feature vectors of unlabeled samples This is the data source for updating the evolutionary prototype. However, to prevent low-quality or erroneous pseudo-labels from contaminating the prototype, these features must first pass through a multi-stage sample quality filter. .like Figure 2 As shown, the detailed workflow of this filter is as follows: Step 1: High-confidence screening. First, based on the model, the unlabeled samples after weak enhancement are screened. The model predicts and filters out samples whose highest predicted probability exceeds a preset threshold (e.g., 0.95). This is an initial coarse screening, retaining samples that the model considers credible.

[0036] ; in, This represents the feature vector of an unlabeled sample that has undergone weak enhancement. This represents a custom activation function (explained in detail in step four) used to process the normalized feature vectors and enhance their expressive power. This represents the transpose of the sparse fusion prototype, where each row of this matrix represents a prototype vector for a class. After transpose, each column represents a prototype. This represents the preset confidence threshold, with a value of 0.95.

[0037] Step 2: Define the consistency boundary. For the sample group that passed the first step screening and belongs to the same pseudo-label category, calculate the geometric center of their feature vectors, i.e., the category center. Then, a dynamic threshold boundary is defined around the center of the category, which is typically determined based on the mean and standard deviation of the distances from the center to all samples within the group.

[0038] Step 3: Outlier filtering. Finally, iterate through all high-confidence samples in this category again, calculating the feature vector of each sample to the category center. The distance is considered. Only samples whose distance is less than the threshold boundary, i.e., those falling within the "consistency" range, are considered truly high-quality samples that are consistent with the characteristics of the class group.

[0039] ; The above formula describes the condition for retaining a sample: the feature vector of the unlabeled sample. Go to its pseudo-label category center The distance must be less than a dynamically calculated threshold; where, Indicates pseudo-tags The category feature center is the center of all pseudo-labels in the current batch. The average value of high-confidence sample features; dynamically calculated distance threshold. It is not a fixed constant, but is determined based on the distance distribution (mean and standard deviation) of samples of that category within the current batch.

[0040] Outliers that are too far away are removed, specifically defined as those where the distance between the sample's feature and the center of its pseudo-label class is greater than (the average distance within the class + 2.0). (Intra-category distance standard deviation). In addition to class consistency checks, this filter can also include feature significance checks (filtering out ambiguous samples with excessively low L2 norms for feature vectors) and prediction certainty checks (filtering out "hesitant" samples where the difference between the highest and second-highest predicted probabilities is too small). Only high-quality features that pass all the above screening steps can be used to update the evolutionary prototype. The update process is conditional, typically requiring a certain number of candidate samples for a category to be triggered, and employs a weighted average method to ensure the prudence of the update.

[0041] Step 3: Classifier construction.

[0042] In each training step, an immediate final classifier for classification needs to be built. First, the anchor prototype from step two... and evolutionary prototype Dynamic weighted fusion is performed to obtain a fusion prototype. .

[0043] ; The formula for calculating the weight β is: ; in, This indicates the current training iteration number. This represents the total number of training iterations. It is 0.99. It is 0.8.

[0044] It is a fusion prototype that changes as the training process progresses, for example, trusting stability more in the early stages of training. gradually increased in the later stages The weights. After fusion, a dense fused prototype is obtained. Subsequently, as... Figure 4 As shown, a sparsification operator This operator is applied to the densely fused prototype. It performs a Top-K selection operation on the prototype vector for each category, retaining only the prototypes with the highest response values ​​(numerical magnitudes). One dimension is set, and the values ​​of the remaining dimensions are set to zero. After this step, the sparse fusion prototype used as the classifier is obtained. .

[0045] ; This sparsity design forces the model to focus on the most discriminative key features, improving the model's robustness and fine-grained resolution.

[0046] Step 4: Classification and Loss Calculation.

[0047] like Figure 1 The lower half Figure 4 and Figure 5 As shown, this step is the core of model parameter optimization.

[0048] Classification matching: First, the feature vector z (containing , , (These are the feature vectors of labeled samples, the feature vectors of unlabeled samples after weak enhancement, and the feature vectors of unlabeled samples after strong enhancement, respectively) are processed by an activation function. Activation function It is a complex processing module that transforms the original feature vector into a sparse, discriminative, and dynamically controlled activated feature vector through gating, scaling, and Leaky Tanh activation, preparing it for subsequent classification and matching. The general steps are as follows: 1. First, find the largest value in the feature vector z that is the “length_threshold” (e.g., the 50th largest value), and subtract this value from the entire vector as the threshold.

[0049] 2. Scale the vector processed in the previous step so that its new maximum value is exactly equal to the hyperparameter M (default value is 3).

[0050] 3. Divide the vector by a "temperature" parameter (default value is 1.5).

[0051] 4. Multiply the final result by a scaling factor (default value 2.0). This is processed into activated features z′. This activation function aims to enhance the discriminative power of the features. Subsequently, the activated features z′ are combined with the sparse fusion prototype constructed in step 3. Perform dot product This process replaces the traditional linear classification layer, obtaining classification prediction results by calculating the similarity between features and prototypes of each category. .

[0052] ; Loss calculation: based on the obtained Calculate three different loss terms: Monitoring loss ( Based on labeled data The predicted results and their true labels y are used to calculate the standard cross-entropy loss.

[0053] ; in, A batch of labeled data represents a single data set containing multiple sample pairs. A set; The cross-entropy loss function measures the difference between two probability distributions. For the model to input samples The output is the original predicted score.

[0054] Unsupervised loss ( Based on unlabeled samples after weak enhancement The prediction results are used to generate pseudo-labels for the samples that pass the high-confidence screening. Then, using the strongly enhanced unlabeled samples... The prediction results and the pseudo-label Calculate the cross-entropy loss. This follows the idea of ​​consistency regularization.

[0055] ; in, This represents a batch of unlabeled data; 'I' is an indicator function, with a value of 1 when the condition within the parentheses is true and 0 when it is false. It is used here for high-confidence filtering. This indicates that the model applies to the m-th unlabeled sample after weak augmentation. The predicted probability distribution; This indicates that the model applies to the m-th strongly augmented unlabeled sample. The predicted probability distribution.

[0056] Agency loss ( ):like Figure 1 The lower left part and Figure 4 As shown, this loss is a key regularization term. It uses the features before activation (feature vectors of labeled samples). and high-quality unlabeled feature vectors ) and sparse fusion prototype By calculating the mean squared error (MSE) between them, the alignment of feature vectors to their corresponding target prototypes is directly enforced in the feature space. This loss term is responsible for "sculpting" a well-structured feature space.

[0057] ; in, This represents a batch of unlabeled data. Let be the feature vector of the i-th labeled sample; Indicates the corresponding real label sparse fusion prototype; This refers to a set of high-quality unlabeled samples, specifically those that have passed a high-confidence screening. This represents the feature vector of the nth high-quality unlabeled sample (from weakly augmented data); Indicates the corresponding pseudo tag A sparse fusion prototype.

[0058] Step 5: Model optimization.

[0059] Finally, the above three types of losses , , We perform a weighted summation to obtain the total loss. The feature extraction network is updated using the gradient of the total loss through backpropagation. The learnable parameter θ is obtained to complete one training iteration.

[0060] ; Among them, unsupervised loss weight The default value is 1, which is the weight of the proxy loss. The default value is 0.8.

[0061] The effectiveness of this invention has been fully validated on the standard semi-supervised learning benchmark dataset CIFAR-10, and its performance comparison results are as follows: Figure 6 As shown. To establish a fair and comparable benchmark, all comparison methods (including advanced models such as FixMatch, FlexMatch, and FullMatch) and the proposed method (ASPEMatch) were run under a unified experimental framework. Key general hyperparameters (such as learning rate, optimizer, batch size, and training epochs) were kept consistent: the optimizer uniformly adopted stochastic gradient descent with momentum (SGD, where the momentum parameter is 0.9), the initial learning rate was set to 0.03 and followed cosine annealing scheduling, the weight decay was 0.0005, the total batch size was 64, and 1,048,576 iterations of training were performed. Meanwhile, the hyperparameters specific to each method were also optimized to the same extent based on their original implementations. Experimental results show that as the number of labeled samples decreased sharply from 4000 to 250, and even 40, the performance of all methods decreased. However, the performance curve of the proposed method showed the most gradual decline, exhibiting the best robustness. Especially in extremely scarce scenarios with only 40 labels, the method of this invention achieved the highest Top-1 accuracy, significantly outperforming other comparative methods, which fully verifies the superiority of this invention in scenarios with few labels.

[0062] Example 2 Semi-supervised image classification systems include: The data input module is used to acquire labeled and unlabeled images; The feature extraction unit, used to extract feature vectors from an image, includes the following units: A dual-track prototype evolution unit is used to maintain and update the anchor prototype and the evolution prototype; The classification and loss calculation unit is used to construct the sparse fusion prototype and perform classification and loss calculation. The model optimization unit is used to optimize the model parameters based on the total loss.

[0063] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0064] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0065] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0066] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0067] Therefore, this invention employs the aforementioned semi-supervised image classification method, system, computer equipment, and medium to address issues such as performance degradation in image classification under few-label scenarios, insufficient robustness to noise and abnormal samples, and lack of interpretability in the model's decision-making process, thereby improving the model's classification accuracy and reliability in real-world complex environments.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A semi-supervised image classification method, characterized in that, Includes the following steps: Acquire image data, including labeled and unlabeled images; The image data is input into the image classification model to obtain the final classification result; The image classification model classifies images by constructing and evolving a dual-track category prototype system, including: The feature extraction module is used to extract feature vectors from the input image; The dual-track prototype evolution module is used to manage and update the anchor prototype and the evolving prototype; The classification and loss calculation module is used to fuse and sparsify the anchored prototype and the evolutionary prototype to generate a sparse fused prototype as a classifier, and calculate the loss based on the classifier to optimize the model. The dual-track prototype evolution module updates the anchor prototype and the evolving prototype in the following ways: The anchor prototype is updated using only the feature vectors of labeled image data via an exponential moving average. The evolutionary prototype is updated using feature vectors from high-quality unlabeled image data filtered through a multi-stage sample quality filter.

2. The semi-supervised image classification method according to claim 1, characterized in that, Multi-stage sample quality filters include: High confidence screening steps: Screen unlabeled samples whose predicted probability is higher than a preset threshold; Category consistency check steps: Calculate the feature center of the unlabeled sample after screening within its pseudo-label category, and remove outlier samples that are too far from the center; An outlier is a sample whose distance to the center of its pseudo-label class is greater than the sum of the average distance of that class and twice the standard deviation of the distance.

3. The semi-supervised image classification method according to claim 1, characterized in that, The classification process in the classification and loss calculation module includes: The feature vector is input into the activation function to obtain the activated feature vector; The anchored prototype and the evolving prototype are dynamically weighted and fused into a dense fused prototype; The classification prediction result is obtained by calculating the dot product between the activated feature vector and the sparse fusion prototype.

4. The semi-supervised image classification method according to claim 3, characterized in that, The loss calculation in the classification and loss calculation module includes: Supervised loss: calculated based on the prediction results and the ground truth labels of labeled images; Unsupervised loss: Pseudo-labels are generated based on predictions of weakly enhanced unlabeled images, and consistency loss is calculated between the prediction results of strongly enhanced unlabeled images; Proxy loss: calculated based on the mean square error between the feature vector and the corresponding target prototype in the sparse fusion prototype.

5. The semi-supervised image classification method according to claim 4, characterized in that, The objects included in the calculation of agency loss are: Feature vectors of labeled images; Feature vectors of high-quality unlabeled image data after being filtered through a multi-stage sample quality filter.

6. The semi-supervised image classification method according to claim 4, characterized in that, The optimization process of an image classification model includes: The total loss is obtained by weighted summing of the supervised loss, the unsupervised loss, and the agency loss. The feature extraction module is optimized by backpropagation based on the total loss.

7. A semi-supervised image classification system, characterized in that, A method for implementing the semi-supervised image classification method as described in any one of claims 1-6, comprising: The data input module is used to acquire labeled and unlabeled images; The feature extraction unit, used to extract feature vectors from an image, includes the following units: A dual-track prototype evolution unit is used to maintain and update the anchor prototype and the evolution prototype; The classification and loss calculation unit is used to construct the sparse fusion prototype and perform classification and loss calculation. The model optimization unit is used to optimize the model parameters based on the total loss.

8. A computer device, characterized in that, It includes a memory and a processor, the memory being used to store instructions and the processor being used to execute the instructions to implement the semi-supervised image classification method as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the semi-supervised image classification method as described in any one of claims 1 to 6.

Images