Image classification method and system based on discriminative atom embedding semi-supervised network
By introducing discriminant atomic and relational dual structures into the image classification method, combined with semi-supervised learning, the problems of insufficient feature distinction and limited labeled data in the prior art are solved, and more efficient image classification performance is achieved.
Patent Information
- Application Number
- CN202210667551.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-14
- Publication Date
- 2025-05-16
- Estimated Expiration
- 2042-06-14
AI Technical Summary
Existing image classification methods are difficult to learn differentiated information when dealing with tasks with small differences between classes and large differences within classes. In the case of limited labeled data, it is difficult for the model to learn enough knowledge, resulting in a decrease in classification accuracy and generalization ability.
A method of image classification based on discriminant atom embedded in semi-supervised network is proposed. Using the classic semi-supervised learning network mean-teacher as the basic framework, it introduces discriminant atom learning distinctive features, constructs a relational dual structure, and improves the distinction of features through discriminant loss function.
By introducing discriminant atomic and relational dual structures, the representation and distinction of features are improved, the problem of low classification accuracy caused by the lack of feature distinction is solved, and the generalization ability of the model is improved through semi-supervised learning when the labeled data is limited.
Smart Images
Figure CN115049876B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of image classification systems, and in particular relates to an image classification method and system based on discriminative atom embedding in a semi-supervised network. Background Art
[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.
[0003] Image classification is an important field of computer vision and has a wide range of application scenarios, such as identity authentication, automatic disease diagnosis, target recognition, etc. Existing image classification methods can achieve good performance to a certain extent. According to the size of the model complexity, they can be divided into the following two methods:
[0004] Methods based on traditional shallow machine learning:
[0005] Traditional machine learning classification methods usually include feature extraction modules and classifier modules. The feature extraction module can use algorithms such as Principal Component Analysis (PCA), Oriented Gradient Histogram (HOG), Scale-Invariant Feature Transform (SIFT) to extract useful feature information from texture features such as edge, color, and size of the target area. The classifier module can use Bayesian Classifier, Random Forest, Support Vector Machine (SVM), etc. for image classification. However, these methods often need to rely on prior information inherent in the task, resulting in poor model robustness, and the performance of the algorithm depends on the feature information provided by the doctor.
[0006] Deep learning based methods:
[0007] In recent years, deep learning has made significant breakthroughs in image recognition and classification tasks. In view of this, researchers have applied deep learning to the field of image classification. The more classic deep classification network models include VGG, Inception V3, Resnet, Densenet, etc. This type of method has become the mainstream image classification method because of its complex structure and can learn more accurate feature representation based on big data.
[0008] Compared with traditional shallow methods, deep learning models can achieve significant improvements, but they still face great challenges in some tasks, which limits the improvement of classification performance. On the one hand, some tasks have certain characteristics of small differences between classes and large differences within classes. Existing deep learning methods are difficult to learn discriminative information, which reduces classification accuracy. On the other hand, the labeled data in some tasks is very limited (such as medical image processing), making it difficult for the model to learn enough knowledge, reducing the generalization ability of the model. Summary of the invention
[0009] In order to solve the above problems, the present invention proposes an image classification method and system based on embedding discriminative atoms into a semi-supervised network. Considering that semi-supervised learning is an effective way to solve the problem of limited labeled data, the present invention uses the classic semi-supervised learning network mean-teacher as the basic framework, introduces discriminative atoms to learn discriminative features, and constructs a relational dual structure. Each sub-network includes a feature extractor and a relational learner. Among them, the relational learner can learn the relationship features between the input image features and the discriminative atomic features, and improve the representation ability of the features. In order to improve the discriminability of the features, a discriminative loss function is introduced, so that the discriminative atoms have stronger discriminability, thereby improving the discriminability of the learned features.
[0010] According to some embodiments, a first solution of the present invention provides an image classification method based on discriminative atomic embedding semi-supervised network, which adopts the following technical solution:
[0011] Image classification method based on discriminative atomic embedding semi-supervised network, including:
[0012] Get input image data;
[0013] Preprocess the input image data to obtain the initial discriminative atoms of each subclass;
[0014] The pre-trained semi-supervised network image classification model is used to compare and predict the input image and the initialized discriminative atoms to obtain the comparison score of each subclass, and the subclass and class corresponding to the maximum score are calculated as the predicted classification result;
[0015] The training process of the semi-supervised network image classification model includes:
[0016] Obtain a sample image dataset and sample initialized discriminative atoms;
[0017] Randomly add noise to the sample image data set to obtain a noisy sample image data set;
[0018] Input the sample image dataset and sample-initialized discriminative atoms into the student model for classification, and input the noisy sample image dataset and sample-initialized discriminative atoms into the teacher model for classification;
[0019] The total loss is obtained based on the classification prediction probability, and training is continuously performed to make the total loss show a downward trend until the training round reaches the set value or the total loss shows a stable trend. The network model with the minimum loss value is saved to obtain the trained semi-supervised network image classification model.
[0020] Furthermore, the process of obtaining the initialized discriminative atom is as follows:
[0021] Obtain image dataset for hierarchical clustering;
[0022] Divide the clustered image dataset into K subclasses;
[0023] Get the cluster center of each subclass as the initial discriminative atom of each subclass.
[0024] Furthermore, the sample image dataset includes a labeled sample image dataset and an unlabeled sample image dataset;
[0025] The noisy sample image dataset includes a labeled noisy sample image dataset and an unlabeled noisy sample image dataset.
[0026] Furthermore, the classification process of the student model includes:
[0027] Perform feature extraction on the sample image data set and the sample initialization discriminant atom to obtain the sample image data set vector and the discriminant atom vector;
[0028] Perform discriminative learning on sample image dataset vectors and discriminative atomic vectors;
[0029] Perform feature concatenation on the sample image dataset vector and the discriminative atomic vector to obtain the concatenated total vector feature;
[0030] The concatenated total vector features are input into the relational learner to obtain the relation score between the sample image dataset vector and the discriminative atomic vector, that is, the predicted probability of each subclass of the sample image.
[0031] Furthermore, the classification process of the teacher model includes:
[0032] Perform feature extraction on the noisy sample image data set and the sample initialization discriminant atom to obtain the noisy sample image data set vector and the discriminant atom vector;
[0033] The noisy sample image dataset vector and the discriminative atomic vector are feature concatenated to obtain the concatenated total vector feature;
[0034] The concatenated total vector features are input into the relational learner to obtain the relation scores between the noisy sample image dataset and the discriminative atomic vector, that is, the predicted probability of each subclass of the noisy sample image.
[0035] Furthermore, the predicted probability of labeled sample image data in the student model is used to calculate the cross entropy loss between it and the true label category;
[0036] The subclass prediction probability of the unlabeled sample image data in the student model and the subclass prediction probability of the unlabeled noisy sample image data in the teacher model are used to calculate the fine-grained consistency loss between the two.
[0037] Using the prediction probability of the unlabeled sample image data in the student model and the prediction probability of the unlabeled noisy sample image data in the teacher model, calculate the label consistency loss of the two;
[0038] Using the labeled sample image dataset vector and the discriminative atomic vector in the student model, calculate the discriminative loss of the two;
[0039] The total loss is obtained by summing the cross entropy loss, the weighted fine-grained consistency loss, the weighted labeling consistency loss, and the weighted discriminability loss.
[0040] Furthermore, the feature extraction module for feature extraction is composed of a shallow convolutional neural network with four convolutional layers and two maximum pooling layers;
[0041] The relation learner consists of two convolution blocks, two maximum pooling layers and two fully connected layers.
[0042] According to some embodiments, a second solution of the present invention provides an image classification system based on discriminative atomic embedding semi-supervised network, which adopts the following technical solution:
[0043] Image classification system based on discriminative atomic embedding semi-supervised network, including:
[0044] A data acquisition module is configured to acquire input image data;
[0045] A data processing module is configured to perform preprocessing based on the input image data to obtain the initialization discriminative atoms of each subclass;
[0046] The image classification module is configured to use a pre-trained semi-supervised network image classification model to compare and predict the input image and the initialized discriminative atom to obtain a comparison score for each subclass, and calculate the subclass and class corresponding to the maximum score as the predicted classification result;
[0047] The training process of the semi-supervised network image classification model includes:
[0048] Obtain a sample image dataset and sample initialized discriminative atoms;
[0049] Randomly add noise to the sample image data set to obtain a noisy sample image data set;
[0050] Input the sample image dataset and sample-initialized discriminative atoms into the student model for classification, and input the noisy sample image dataset and sample-initialized discriminative atoms into the teacher model for classification;
[0051] The total loss is obtained based on the classification prediction probability, and training is continuously performed to make the total loss show a downward trend until the training round reaches the set value or the total loss shows a stable trend. The network model with the minimum loss value is saved to obtain the trained semi-supervised network image classification model.
[0052] According to some embodiments, a third aspect of the present invention provides a computer-readable storage medium.
[0053] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the image classification method based on discriminative atom embedding semi-supervised network as described in the first aspect above.
[0054] According to some embodiments, a fourth aspect of the present invention provides a computer device.
[0055] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the image classification method based on discriminative atomic embedding semi-supervised network as described in the first aspect above are implemented.
[0056] Compared with the prior art, the present invention has the following beneficial effects:
[0057] The present invention proposes an image classification method and system based on a semi-supervised network embedded with discriminative atoms, which has excellent performance in image classification compared with previous methods. On the one hand, the present invention cites the idea of discriminative atoms as class templates for each subclass, introduces a dual network to learn the relationship features between input images and atoms, improves the representation of the learned features, introduces discriminative loss, and further improves the discriminativeness of the learned features, solving the problem that the lack of feature discriminativeness is difficult to solve, the similarity between classes is large, and the similarity within classes is small, resulting in low classification accuracy.
[0058] On the other hand, the present invention uses the semi-supervised learning idea to target the difficulty of obtaining labeled data in the field of image classification. It can complete the deep learning training process with only a small amount of labeled data, and has good learning effects. The algorithm of the semi-supervised learning idea of basic discriminative atomic embedding and relational duality can meet the deep learning requirements of image classification tasks. This idea can also be applied to other fields with this problem to help other fields better complete deep learning tasks. BRIEF DESCRIPTION OF THE DRAWINGS
[0059] 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.
[0060] Figure 1 is a flow chart of an image classification method based on discriminative atom embedding semi-supervised network according to an embodiment of the present invention;
[0061] Figure 2 It is a schematic diagram of the framework of the semi-supervised network image classification model described in an embodiment of the present invention. DETAILED DESCRIPTION
[0062] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0063] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.
[0064] 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 when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.
[0065] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.
[0066] Embodiment 1
[0067] like Figure 1-Figure 2 As shown, this embodiment provides an image classification method based on discriminative atomic embedding semi-supervised network, including:
[0068] Get input image data;
[0069] Preprocess the input image data to obtain the initial discriminative atoms of each subclass;
[0070] The pre-trained semi-supervised network image classification model is used to compare and predict the input image and the initialized discriminative atoms to obtain the comparison score of each subclass, and the subclass and class corresponding to the maximum score are calculated as the predicted classification result;
[0071] The training process of the semi-supervised network image classification model includes:
[0072] Obtain a sample image dataset and sample initialized discriminative atoms;
[0073] Randomly add noise to the sample image data set to obtain a noisy sample image data set;
[0074] Input the sample image dataset and sample-initialized discriminative atoms into the student model for classification, and input the noisy sample image dataset and sample-initialized discriminative atoms into the teacher model for classification;
[0075] The total loss is obtained based on the classification prediction probability, and training is continuously performed to make the total loss show a downward trend until the training round reaches the set value or the total loss shows a stable trend. The network model with the minimum loss value is saved to obtain the trained semi-supervised network image classification model.
[0076] like Figure 1 As shown, the method described in this embodiment includes the following steps:
[0077] Step S1: Acquire image data and initialize discriminative atoms, including:
[0078] Step S1.1: First, perform hierarchical clustering on the image data set. Hierarchical clustering is a common unsupervised clustering method that measures similarity by calculating the distance between data points. The two data points with the closest distance are merged to generate a nested clustering tree. The tree diagram can clearly show its hierarchical structure. Commonly used distance measurement methods include: Manhattan distance, Euclidean distance, and Chebyshev distance. The present invention uses a hierarchical clustering algorithm based on Euclidean distance to cluster labeled data sets.
[0079]
[0080] Among them, x and y are points in n-dimensional space.
[0081] Since image categories in some tasks have the characteristics of small inter-class differences and large intra-class differences, in order to make the model easier to distinguish CNV types, the present invention divides the clustered data set into K subclasses and selects relevant images as discriminative atoms for initialization.
[0082] K=mC,m>1&& m∈N + (2)
[0083] Among them, K is the number of subclasses, C is the number of primitive data types, and m is any natural number greater than one.
[0084] Each category is divided into m subcategories, which ensures the category balance and improves the distinguishability of features, making it easier for the model to distinguish subcategories.
[0085] Step S1.2: Discriminative Atom Initialization
[0086] According to the clustering results and partitioning results in step S1.1, a suitable image is selected as the initial discriminative atom. The discriminative atom is the class center of the subclass and represents the characteristics of the subclass. The initial discriminative atom is obtained from the cluster center of the hierarchical clustering. The cluster center of each subclass represents the discriminative atom of each subclass.
[0087] In addition, the image data set is randomly noised to obtain a noisy image data set;
[0088] The image dataset includes a labeled image dataset and an unlabeled image dataset;
[0089] The noisy image dataset includes a labeled noisy image dataset and an unlabeled noisy image dataset.
[0090] Step S2: Use the pre-trained semi-supervised network image classification model to predict and obtain the comparison score of each category, and calculate the category corresponding to the maximum score as the predicted classification result, including:
[0091] Step S2.1: Feature extraction
[0092] For the image data obtained in step S1, the key feature information vector of the image is extracted through the feature extraction module f(x, φ) (where x is the input vector and φ is the model parameter) to obtain the image data set vector and the discriminant atomic vector (where K represents the number of subcategories).
[0093] The labeled image dataset vector is (where n represents the number of labeled data), the unlabeled image dataset vector is (where Nn represents the number of unlabeled data).
[0094] The feature extraction module is composed of a shallow convolutional neural network with four convolutional layers and two maximum pooling layers. The feature vectors are extracted by the same network model and therefore belong to the same feature space. The Student model and the Teacher model are run simultaneously in this process, with the same feature extraction module and different input data. The image input to the Student model is the original image, and the image input to the Teacher model is the image after random noise addition.
[0095] The labeled data and unlabeled data are randomly denoised, and the noise includes four random combinations of displacement, image brightness, contrast, and saturation. The displacement value, image brightness, contrast, and saturation change values are all random numbers within a certain range.
[0096] Step S2.2: Feature stitching
[0097] The data set vector X obtained in S2.1 l , X u The student model and the teacher model in this step are the same.
[0098] Step S2.3: Discriminative Learning
[0099] The input image and the initialized discriminative atoms are passed through the network's feature extraction module for discriminative learning. The discriminative atoms are dynamically adjusted to the feature vector center of the corresponding labeled subclass, and the discriminative loss is used to shorten the distance between the labeled sample and the corresponding discriminative atoms, further enhancing the discriminativeness of the features. The discriminative atoms are an important basis for the learning and classification of the relational learner, and directly affect the performance of the model. This step is only performed on the Student model.
[0100] Step S2.4: Relationship Learning
[0101] The relational learner consists of two convolutional blocks, two maximum pooling layers, and two fully connected layers. The total vector after feature concatenation is input into the relational learner g(y,φ), and the relationship score between the vector of the input image and each discriminant atomic vector can be obtained, that is, the predicted probability of each subcategory of the input image is obtained. Given the relationship between subcategories and categories, the predicted probability of the category can be derived from the predicted probability of the subcategory. In this step, the student model and the teacher model are the same. The output prediction probability includes not only the results of labeled data but also the results of unlabeled data.
[0102] Step S3: Obtain the total loss of the semi-supervised network image classification model, including:
[0103] Step S3.1: Calculate cross entropy loss
[0104] Use the labeled data X of all Student modules l Predict the probability and calculate its difference with the true label category y l Perform cross entropy loss.
[0105]
[0106] Among them, y i is x i The true label of , θ is the parameter of the Student model, f θ is the Student model. This classification loss can effectively ensure that the network learns the correct parameters and constrains the network using labeled data.
[0107] Step S3.2: Calculate fine-grained consistency loss
[0108] Use all the unlabeled data X of the Student module u The subclass prediction probability and the unlabeled data X of all Teacher modules u The subclass prediction probability of is calculated and the consistency loss between the two is calculated.
[0109]
[0110] in, is the network model that outputs the subclass results, θ' is the parameter of the Teacher model, and η and η' are random noises. Fine-grained consistency loss is a major innovation of this invention. By making the subclass results predicted by the unlabeled data output by the two branches of the network consistent as much as possible, the network can be strongly constrained to mine useful information from the unlabeled data.
[0111] Step S3.3: Calculate label consistency loss
[0112] Use all the unlabeled data X of the Student module u The predicted probability and the unlabeled data X of all Teacher modules u The predicted probability of , and calculate the consistency loss between the two.
[0113]
[0114] The label consistency loss constrains the classification results output by the two branches of the network, which can weakly constrain the network to obtain information from unlabeled data.
[0115] Step S3.4: Calculate the discriminative loss
[0116] Using the labeled data X of the Student module l and the discriminative atomic vector V, and calculate the loss of both.
[0117]
[0118] in, is x i The corresponding subclass label category, yes The discriminative atom of the corresponding category. The discriminative loss is mainly used to constrain the discriminative learning module of the Student branch in the network and enhance the discriminativeness.
[0119] Step S3.5: Joint training
[0120] The cross entropy loss Loss c , Fine-grained consistency loss Loss FJ Weighted, labeled consistency loss Loss J Weighted and discriminative loss v Weighted together as the total loss Loss = Loss c +λLoss FJ +ωLoss J +μLoss v (where λ, ω, and μ are hyperparameters), and training is continuously performed to make the Loss show a downward trend until the training round reaches the set value or the Loss shows a stable trend. The network model with the minimum loss value is saved as the trained semi-supervised network image classification model.
[0121] like Figure 1 As shown in the figure, the system in the dotted box is the system module that mainly performs the classification function, in which the feature vector module uses the network model f(x, φ) described in step S2.1, and the comparison score acquisition module uses the network model g(y, φ) described in step S2.4. They have determined the appropriate network parameters φ, φ through the training process.
[0122] The user inputs the image data to be tested into the classification system, and the classification system automatically performs three processes: feature vector extraction, comparison score acquisition, and prediction category calculation. Finally, the prediction category is output to interact with the user.
[0123] Embodiment 2
[0124] This embodiment provides an image classification system based on discriminative atomic embedding semi-supervised network.
[0125] A data acquisition module is configured to acquire input image data;
[0126] A data processing module is configured to perform preprocessing based on the input image data to obtain the initialization discriminative atoms of each subclass;
[0127] The image classification module is configured to use a pre-trained semi-supervised network image classification model to compare and predict the input image and the initialized discriminative atom to obtain a comparison score for each subclass, and calculate the subclass and class corresponding to the maximum score as the predicted classification result;
[0128] The training process of the semi-supervised network image classification model includes:
[0129] Obtain a sample image dataset and sample initialized discriminative atoms;
[0130] Randomly add noise to the sample image data set to obtain a noisy sample image data set;
[0131] Input the sample image dataset and sample-initialized discriminative atoms into the student model for classification, and input the noisy sample image dataset and sample-initialized discriminative atoms into the teacher model for classification;
[0132] The total loss is obtained based on the classification prediction probability, and training is continuously performed to make the total loss show a downward trend until the training round reaches the set value or the total loss shows a stable trend. The network model with the minimum loss value is saved to obtain the trained semi-supervised network image classification model.
[0133] The examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the contents disclosed in the above embodiment 1. It should be noted that the above modules as part of the system can be executed in a computer system such as a set of computer executable instructions.
[0134] 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.
[0135] The proposed system can be implemented in other ways. For example, the system embodiment described above is only illustrative, and the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.
[0136] Embodiment 3
[0137] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the image classification method based on discriminative atom embedding semi-supervised network as described in the first embodiment are implemented.
[0138] Embodiment 4
[0139] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the image classification method based on discriminative atom embedding semi-supervised network as described in the first embodiment are implemented.
[0140] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.
[0141] 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.
[0142] 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.
[0143] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0144] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.
[0145] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.
Claims
1. An image classification method based on discriminative atomic embedding semi-supervised network, characterized in that: include: Get input image data; Preprocessing is performed based on the input image data to obtain the initial discriminant atom of each subclass; wherein the process of obtaining the initial discriminant atom is as follows: obtaining the image data set for hierarchical clustering; dividing the clustered image data set into K subclasses; obtaining the cluster center of each subclass as the initial discriminant atom of each subclass; The pre-trained semi-supervised network image classification model is used to compare and predict the input image and the initialized discriminative atoms to obtain the comparison score of each subclass, and the subclass and class corresponding to the maximum score are calculated as the predicted classification result; The training process of the semi-supervised network image classification model includes: Obtain a sample image dataset and sample initialized discriminative atoms; Randomly add noise to the sample image data set to obtain a noisy sample image data set; The sample image data set and the sample initialized discriminative atoms are input into the student model for classification. The classification process of the student model includes: extracting features from the sample image data set and the sample initialized discriminative atoms to obtain the sample image data set vector and the discriminative atom vector; performing discriminative learning on the sample image data set vector and the discriminative atom vector; performing feature concatenation on the sample image data set vector and the discriminative atom vector to obtain the concatenated total vector feature; inputting the concatenated total vector feature into the relational learner to obtain the relation score between the sample image data set vector and the discriminative atom vector, and using the relation score as the predicted probability of each subclass of the sample image; Inputting the noisy sample image data set and the sample initialized discriminative atoms into the teacher model for classification; the classification process of the teacher model includes: extracting features from the noisy sample image data set and the sample initialized discriminative atoms to obtain the noisy sample image data set vector and the discriminative atom vector; performing feature concatenation on the noisy sample image data set vector and the discriminative atom vector to obtain the concatenated total vector feature; inputting the concatenated total vector feature into the relational learner to obtain the relation score between the noisy sample image data set vector and the discriminative atom vector, and using the relation score as the predicted probability of each subclass of the noisy sample image; The total loss is obtained based on the classification prediction probability, and training is continued to make the total loss show a downward trend until the training round reaches the set value or the total loss shows a stable trend. The network model with the minimum loss value is saved to obtain the trained semi-supervised network image classification model; The process of obtaining the total loss is as follows: using the predicted probability of the labeled sample image data in the student model, calculate the cross entropy loss with the true label category; using the subclass prediction probability of the unlabeled sample image data in the student model and the subclass prediction probability of the unlabeled noisy sample image data in the teacher model, calculate the fine-grained consistency loss of the two; using the predicted probability of the unlabeled sample image data in the student model and the predicted probability of the unlabeled noisy sample image data in the teacher model, calculate the label consistency loss of the two; using the labeled sample image data set vector and the discriminative atomic vector in the student model, calculate the discriminative loss of the two; weighted summing the cross entropy loss, the fine-grained consistency loss, the label consistency loss and the discriminative loss to obtain the total loss.
2. The image classification method based on discriminative atomic embedding semi-supervised network as claimed in claim 1, characterized in that: The sample image dataset includes a labeled sample image dataset and an unlabeled sample image dataset; The noisy sample image dataset includes a labeled noisy sample image dataset and an unlabeled noisy sample image dataset.
3. The image classification method based on discriminative atomic embedding semi-supervised network as claimed in claim 1, characterized in that: The feature extraction module for feature extraction is composed of a shallow convolutional neural network with four convolutional layers and two maximum pooling layers; The relation learner consists of two convolution blocks, two maximum pooling layers and two fully connected layers.
4. Image classification system based on discriminative atomic embedding semi-supervised network, characterized by: include: A data acquisition module is configured to acquire input image data; The data processing module is configured to perform preprocessing based on the input image data to obtain the initial discriminant atom of each subclass; wherein the process of obtaining the initial discriminant atom is as follows: obtaining the image data set for hierarchical clustering; dividing the clustered image data set into K subclasses; obtaining the cluster center of each subclass as the initial discriminant atom of each subclass; The image classification module is configured to use a pre-trained semi-supervised network image classification model to compare and predict the input image and the initialized discriminative atom to obtain a comparison score for each subclass, and calculate the subclass and class corresponding to the maximum score as the predicted classification result; The training process of the semi-supervised network image classification model includes: Obtain a sample image dataset and sample initialized discriminative atoms; Randomly add noise to the sample image data set to obtain a noisy sample image data set; The sample image data set and the sample initialized discriminative atoms are input into the student model for classification. The classification process of the student model includes: extracting features from the sample image data set and the sample initialized discriminative atoms to obtain the sample image data set vector and the discriminative atom vector; performing discriminative learning on the sample image data set vector and the discriminative atom vector; performing feature concatenation on the sample image data set vector and the discriminative atom vector to obtain the concatenated total vector feature; inputting the concatenated total vector feature into the relational learner to obtain the relation score between the sample image data set vector and the discriminative atom vector, and using the relation score as the predicted probability of each subclass of the sample image; Inputting the noisy sample image data set and the sample initialized discriminative atoms into the teacher model for classification; the classification process of the teacher model includes: extracting features from the noisy sample image data set and the sample initialized discriminative atoms to obtain the noisy sample image data set vector and the discriminative atom vector; performing feature concatenation on the noisy sample image data set vector and the discriminative atom vector to obtain the concatenated total vector feature; inputting the concatenated total vector feature into the relational learner to obtain the relation score between the noisy sample image data set vector and the discriminative atom vector, and using the relation score as the predicted probability of each subclass of the noisy sample image; The total loss is obtained based on the classification prediction probability, and training is continued to make the total loss show a downward trend until the training round reaches the set value or the total loss shows a stable trend. The network model with the minimum loss value is saved to obtain the trained semi-supervised network image classification model; The process of obtaining the total loss is as follows: using the predicted probability of the labeled sample image data in the student model, calculate the cross entropy loss with the true label category; using the subclass prediction probability of the unlabeled sample image data in the student model and the subclass prediction probability of the unlabeled noisy sample image data in the teacher model, calculate the fine-grained consistency loss of the two; using the predicted probability of the unlabeled sample image data in the student model and the predicted probability of the unlabeled noisy sample image data in the teacher model, calculate the label consistency loss of the two; using the labeled sample image data set vector and the discriminative atomic vector in the student model, calculate the discriminative loss of the two; weighted summing the cross entropy loss, the fine-grained consistency loss, the label consistency loss and the discriminative loss to obtain the total loss.
5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the image classification method based on discriminative atom embedding semi-supervised network as described in any one of claims 1 to 3 are implemented.
6. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the image classification method based on discriminative atom embedding semi-supervised network as described in any one of claims 1 to 3 are implemented.