Few-shot image recognition method and device, computer device, and storage medium
By extracting and transforming local features through a backbone network, constructing a weighted similarity score matrix, and optimizing the backbone network, the problems of global feature interference and insufficient annotation in few-sample image recognition are solved, significantly improving recognition accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENZHEN INST OF ADVANCED TECH CHINESE ACAD OF SCI
- Filing Date
- 2023-10-10
- Publication Date
- 2026-05-15
AI Technical Summary
In existing few-shot image recognition technologies, global features and category-independent features interfere with similarity measurement, resulting in low recognition accuracy, and insufficient labeled images affect model performance.
Local features are extracted through a backbone network, and category-independent features are converted into category-related features using a feature propagation mechanism. A densely connected similarity score matrix is constructed and weighted to optimize the backbone network for recognition.
It effectively filters out category-irrelevant features, increases the number of labeled samples, improves the accuracy of recognition of images with few samples, and alleviates the problem of scarce labeled images.
Smart Images
Figure CN117274710B_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of deep learning technology, and specifically relates to a method, apparatus, computer device, and storage medium for few-shot image recognition. Background Technology
[0002] Single-label datasets are commonly used for few-shot image recognition. Since a single label can only describe a portion of the image region, few-shot image recognition faces the following challenges: (1) Because global pooling is the main method for obtaining global image features, class-independent features will inevitably exist in the global image features, leading to inaccurate similarity measurements and low robustness; (2) Due to the extremely small number of labeled images, class-related entities are not dominant in number. Considering the complex variations in scale, shape, and background of class-related entities, class-independent features may even dominate the similarity measurement. These challenges severely restrict the performance of few-shot image recognition.
[0003] Current research on few-shot image recognition mainly focuses on three aspects: data augmentation, meta-learning, and metric learning, as detailed below:
[0004] (1) Data augmentation is to increase the number of samples by using sample generation algorithms or unlabeled samples, thereby solving the problem of insufficient labeled images in the recognition of few-sample images.
[0005] (2) Meta-learning refers to using learned meta-knowledge to guide the model in learning new few-shot image recognition tasks. Meta-knowledge can be model hyperparameters, network initial parameters, network structure, optimizer, etc.
[0006] (3) Metric learning-based few-shot image recognition methods first map the input image to a transferable feature space, where semantically similar images are close together and semantically different images are far apart; then, nearest neighbor search is used to complete the few-shot image recognition task. The construction of the transferable feature space and the design of similarity measurement criteria are the focus of this type of method.
[0007] However, existing few-shot image recognition technologies still have the following drawbacks:
[0008] (1) In data augmentation, meta-learning and metric learning methods, global features plus linear classifier, Euclidean distance or cosine distance are commonly used similarity measurement frameworks. Considering that global pooling is the main way to obtain global features of an image and that the image contains both class-related entities and class-independent backgrounds, class-independent features will inevitably interfere with similarity measurement. Furthermore, considering the complex changes in scale, shape, and background of class-related entities and the extremely small number of labeled images, class-independent features may even dominate similarity measurement.
[0009] (2) The framework of local features plus semantic alignment can weaken the influence of category-independent features, but background matching, category-related entity matching and category-independent entity matching will still interfere with similarity measurement. Summary of the Invention
[0010] This application provides a method, apparatus, computer device, and storage medium for few-sample image recognition, aiming to at least partially solve one of the aforementioned technical problems in the prior art.
[0011] To address the above problems, this application provides the following technical solution:
[0012] A few-shot image recognition method includes:
[0013] Local features of sample images are extracted using a backbone network;
[0014] The category-independent features in the local features are converted into category-related features using a feature propagation mechanism, and the category-relatedness of the converted local features is quantified to obtain the category-related weights of the local features.
[0015] Construct a dense connection similarity score matrix, and then weight the dense connection similarity score matrix using the category-related weights of the local features;
[0016] The similarity score between two sample images is calculated based on the weighted dense connection similarity score matrix. The backbone network is then optimized based on the similarity score, and the optimized backbone network is used for few-sample image recognition.
[0017] The technical solution adopted in this application embodiment further includes: before extracting local features of the sample image using the backbone network, it further includes:
[0018] The backbone network is pre-trained using mask images of the sample images for modeling.
[0019] The technical solution adopted in this application embodiment further includes: after extracting local features of the sample image using a backbone network, it also includes:
[0020] The extracted local features are treated as graph nodes. An adjacency matrix is constructed using these graph nodes, and the Laplacian matrix of the adjacency matrix is then built. The adjacency matrix is as follows:
[0021]
[0022] Where m and n represent the m-th local feature and the n-th local feature, respectively, σ represents the variance of the local feature magnitude, exp represents the exponential function, and points that are farther away are defined as neighboring points;
[0023] The Laplace matrix of the adjacency matrix is:
[0024]
[0025] Where D is a diagonal matrix.
[0026] The technical solution adopted in this application embodiment further includes: using a feature propagation mechanism to convert category-independent features in the local features into category-related features, and quantifying the category relevance of the converted local features to obtain the category relevance weights of the local features, specifically as follows:
[0027] Based on the Laplacian matrix of the adjacency matrix, a feature propagation mechanism is used to enable each local feature to absorb complementary information from its neighbors, transforming class-independent features in the local features into class-related features:
[0028] X′=PX#
[0029] Where P is the propagation matrix and P = (I - αL) -1 α is a balance factor and α∈(0,1);
[0030] The category-related weights of the local features are defined as follows:
[0031] w=X′c T #
[0032] Where X′ is the category-related feature, and c is the class center of the local feature.
[0033] The technical solution adopted in this application embodiment also includes: the construction of the dense connection similarity score matrix specifically involves:
[0034] Given the class-related features X′1 and X′2 of any two sample images, the dense connection similarity score matrix between X′1 and X′2 is:
[0035] S[m,n]=ψ(X′1[m,:],X′2[n,:])#
[0036] Where ψ(·,·) is the cosine function;
[0037] The specific steps for weighting the densely connected similarity score matrix using category-related weights based on local features are as follows:
[0038] S′=W1⊙S⊙W2#
[0039] Where W1 is the category-related weight matrix of the local features of the first sample image, and W2 is the category-related weight matrix of the local features of the second sample image, and W2[m,:] = w2, where ⊙ represents the multiplication of elements at corresponding positions in the matrix.
[0040] The technical solution adopted in this application embodiment further includes: the step of calculating the similarity score between two sample images based on the weighted dense connection similarity score matrix specifically involves:
[0041]
[0042] The technical solution adopted in this application embodiment further includes: the specific method of using the optimized backbone network for few-sample image recognition is as follows:
[0043] The optimized backbone network is used to obtain the similarity score between the test image and any labeled image, and the label of the labeled image with the highest similarity score is assigned to the test image to generate the recognition result of the test image.
[0044] Another technical solution adopted in this application embodiment is: a few-sample image recognition device, comprising:
[0045] Feature extraction module: used to extract local features of sample images using the backbone network;
[0046] Feature transformation module: used to convert class-independent features in the local features into class-related features using a feature propagation mechanism, and to quantify the class relevance of the transformed local features to obtain the class relevance weights of the local features;
[0047] Similarity score weighting module: used to construct a dense connection similarity score matrix and to weight the dense connection similarity score matrix using the category correlation weights of the local features;
[0048] Model optimization module: used to calculate the similarity score between two sample images based on the weighted dense connection similarity score matrix, optimize the backbone network based on the similarity score, and use the optimized backbone network to perform few-sample image recognition.
[0049] Another technical solution adopted in this application embodiment is: a computer device, the computer device including a processor and a memory coupled to the processor, wherein,
[0050] The memory stores program instructions for implementing the few-sample image recognition method;
[0051] The processor is used to execute the program instructions stored in the memory to control the few-sample image recognition method.
[0052] Another technical solution adopted in this application embodiment is: a storage medium storing processor-executable program instructions, the program instructions being used to execute the few-sample image recognition method.
[0053] Compared with the prior art, the beneficial effects of the embodiments of this application are as follows: The few-sample image recognition method, apparatus, computer device and storage medium of the embodiments of this application utilize the feature propagation mechanism to enable each local feature to absorb complementary information from its neighboring points, thereby converting the category-irrelevant features in the local features into category-related features. This not only achieves the filtering out of category-irrelevant features, but also increases the number of labeled samples, thereby alleviating the problem of scarce labeled images. At the same time, it avoids the problem of difficulty in locating category-related features, and can significantly improve the accuracy of few-sample image recognition. Attached Figure Description
[0054] Figure 1 This is a flowchart of the few-sample image recognition method according to the first embodiment of this application;
[0055] Figure 2 This is a flowchart of the few-sample image recognition method according to the second embodiment of this application;
[0056] Figure 3 This is a schematic diagram of the structure of a few-sample image recognition device according to an embodiment of this application;
[0057] Figure 4 This is a schematic diagram of the computer device structure according to an embodiment of this application;
[0058] Figure 5 This is a schematic diagram of the structure of the storage medium according to an embodiment of this application. Detailed Implementation
[0059] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0060] The terms "first," "second," and "third" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or computer device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or computer devices.
[0061] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.
[0062] Please see Figure 1 This is a flowchart of a few-shot image recognition method according to a first embodiment of this application. The few-shot image recognition method according to the first embodiment of this application includes the following steps:
[0063] S100: Extract local features of sample images using a backbone network;
[0064] S110: The category-independent features in the local features are converted into category-related features using the feature propagation mechanism, and the category-relatedness of the converted local features is quantified to obtain the category-related weights of the local features;
[0065] S120: Construct a dense connection similarity score matrix, and use the category-related weights of the local features to weight the dense connection similarity score matrix;
[0066] S130: Calculate the similarity score between two sample images based on the weighted dense connection similarity score matrix, optimize the backbone network based on the similarity score, and use the optimized backbone network to perform few-sample image recognition.
[0067] Please see Figure 2 This is a flowchart of a few-shot image recognition method according to a second embodiment of this application. The few-shot image recognition method according to the second embodiment of this application includes the following steps:
[0068] S200: Obtain any two sample images and their corresponding mask images from a single-label dataset;
[0069] In this step, the single-label dataset is a commonly used dataset for few-shot image recognition, and the sample images obtained are single-label images.
[0070] S210: The backbone network is pre-trained using mask images of sample images, and local features of the sample images are extracted through the backbone network.
[0071] In this step, mask image modeling is used to train the backbone network, which can achieve a good understanding of the image structure. This application embodiment uses... This represents the local features extracted using the backbone network, where P represents the number of local features and F represents the dimension of the local features.
[0072] S220: Treat the extracted local features as graph nodes, construct an adjacency matrix using the graph nodes, and construct the Laplacian matrix of the adjacency matrix.
[0073] In this step, since a single label can only describe a portion of the image region, not all local features are category-related. Considering the scarcity of labeled images, the fact that category-related entities are not numerous, and the complex variations in the scale, shape, and background of category-related entities, the backbone network faces challenges in accurately locating category-related entities. To avoid the challenges of locating category-related features, this embodiment defines the adjacency matrix as follows:
[0074]
[0075] Where m and n represent the m-th and n-th local features, respectively, σ represents the variance of the local feature magnitude, exp represents the exponential function, and points that are farther away are defined as neighboring points (i.e., neighboring features). Then, the Laplace matrix of the adjacency matrix is defined as:
[0076]
[0077] Where D is a diagonal matrix.
[0078] S230: A Laplacian matrix based on the adjacency matrix, which uses a feature propagation mechanism to enable each local feature to absorb complementary information from its neighbors, transforming class-independent features in the local features into class-related features;
[0079] In this step, class-independent features in the local features are transformed into class-related features through the following feature propagation mechanism:
[0080] X′=PX#(3)
[0081] Where P is the propagation matrix and P = (I - αL) -1 α is a balance factor and α∈(0,1). In the propagation matrix P, larger weights are assigned to neighboring points, thus a local feature absorbs the most information from its neighboring points. Since this embodiment defines distant points as neighboring points, a local feature absorbs the most information from distant local features. It can be understood that this embodiment utilizes a feature propagation mechanism to enable each local feature to absorb complementary information from its neighboring points, converting class-irrelevant features in the local features into class-related features. This achieves the filtering out of class-irrelevant features, increases the number of labeled samples to alleviate the problem of scarce labeled samples, and avoids the difficulty in locating class-related features.
[0082] S240: Quantify the class correlation of the transformed local features to obtain the class correlation weights of the local features;
[0083] In this step, the following problems arise when using feature propagation for feature category conversion: If a local feature is a category-independent feature, it will absorb more information from category-related features and become category-related because they are far away; conversely, if a local feature is a category-related feature, it will absorb more information from category-independent features because they are far away, thus weakening its category relevance. To eliminate these problems, the category relevance weights of local features are defined in this embodiment as follows:
[0084] w=X′c T #(4)
[0085] Where X′ represents the class-related feature, and c represents the class center of the local feature.
[0086] S250: Construct a dense connection similarity score matrix and weight the dense connection similarity score matrix using the category correlation weights of local features;
[0087] In this step, given the category-related features of any two sample images as X′1 and X′2, the dense connection similarity score matrix between X′1 and X′2 is defined in this embodiment as follows:
[0088] S[m,n]=ψ(X′1[m,:],X′2[n,:])#(5)
[0089] Where ψ(·,·) is the cosine function.
[0090] To eliminate the interference of outliers that may exist due to feature transformation on the similarity measure, embodiments of this application utilize the category-related weights of local features to weight the densely connected similarity score matrix:
[0091] S′=W1⊙S⊙W2#(6)
[0092] Where W1 is the category-related weight matrix of the local features of the first sample image, and W2 is the category-related weight matrix of the local features of the second sample image, and W2[m,:] = w2, where ⊙ represents the multiplication of elements at corresponding positions in the matrix.
[0093] S260: Calculate the similarity score between two sample images based on the weighted dense connection similarity score matrix, and construct a loss function based on the similarity score to optimize the backbone network and enhance the class correlation of local features;
[0094] In this step, the similarity score between sample images is calculated as follows:
[0095]
[0096] S270: Using an optimized backbone network for few-shot image recognition;
[0097] In this step, when performing few-sample image recognition, the backbone network is used to obtain the similarity score between the test image and any labeled image, and the label of the labeled image with the highest similarity score is assigned to the test image to generate the recognition result of the test image.
[0098] Based on the above, the few-shot image recognition method of this application utilizes the feature propagation mechanism to enable each local feature to absorb complementary information from its neighboring points, thereby converting class-independent features in the local features into class-related features. This not only achieves the filtering out of class-independent features but also increases the number of labeled samples, thereby alleviating the problem of scarce labeled images. At the same time, it avoids the problem of difficulty in locating class-related features, and can significantly improve the accuracy of few-shot image recognition.
[0099] Please see Figure 3 This is a schematic diagram of the structure of a few-shot image recognition device according to an embodiment of this application. The few-shot image recognition device 40 according to an embodiment of this application includes:
[0100] Feature extraction module 41: used to extract local features of sample images using the backbone network;
[0101] Feature transformation module 42: used to convert class-independent features in the local features into class-related features using a feature propagation mechanism, and to quantify the class-relatedness of the transformed local features to obtain the class-related weights of the local features;
[0102] Similarity score weighting module 43: used to construct a dense connection similarity score matrix and to weight the dense connection similarity score matrix using the category correlation weights of the local features;
[0103] Model optimization module 44: is used to calculate the similarity score between two sample images based on the weighted dense connection similarity score matrix, optimize the backbone network based on the similarity score, and use the optimized backbone network to perform few-sample image recognition.
[0104] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.
[0105] The apparatus provided in this application can be applied to the foregoing method embodiments. For details, please refer to the description of the above method embodiments, which will not be repeated here.
[0106] Please see Figure 4 This is a schematic diagram of a computer device structure according to an embodiment of this application. The computer device 50 includes:
[0107] Memory 51 storing executable program instructions;
[0108] Processor 52 connected to memory 51;
[0109] The processor 52 is used to call the executable program instructions stored in the memory 51 and perform the following steps: extracting local features of the sample images using a backbone network; converting class-independent features in the local features into class-related features using a feature propagation mechanism, and quantifying the class-relatedness of the converted local features to obtain the class-related weights of the local features; constructing a dense connection similarity score matrix, and weighting the dense connection similarity score matrix using the class-related weights of the local features; calculating the similarity score between two sample images based on the weighted dense connection similarity score matrix, optimizing the backbone network based on the similarity score, and using the optimized backbone network for few-sample image recognition.
[0110] The processor 52 can also be referred to as a CPU (Central Processing Unit). The processor 52 may be an integrated circuit chip with signal processing capabilities. The processor 52 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.
[0111] Please see Figure 5 This is a schematic diagram of the structure of the storage medium in an embodiment of this application. The storage medium in this embodiment stores program instructions 61 capable of implementing the following steps: extracting local features of sample images using a backbone network; converting class-independent features in the local features into class-related features using a feature propagation mechanism, and quantifying the class-relatedness of the converted local features to obtain class-related weights of the local features; constructing a dense connection similarity score matrix, and weighting the dense connection similarity score matrix using the class-related weights of the local features; calculating the similarity score between two sample images based on the weighted dense connection similarity score matrix, optimizing the backbone network based on the similarity score, and performing few-sample image recognition using the optimized backbone network. The program instructions 61 can be stored in the aforementioned storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network computer device, etc.) or processor to execute all or part of the steps of the methods in various embodiments of this application. The aforementioned storage media include: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program instructions, or terminal computer devices such as computers, servers, mobile phones, and tablets. Servers can be standalone servers or cloud servers providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.
[0112] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, apparatuses, or units, and may be electrical, mechanical, or other forms.
[0113] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.
Claims
1. A method for few-sample image recognition, characterized in that, include: Local features of sample images are extracted using a backbone network; The category-independent features in the local features are converted into category-related features using a feature propagation mechanism, and the category-relatedness of the converted local features is quantified to obtain the category-related weights of the local features. Construct a dense connection similarity score matrix, and then weight the dense connection similarity score matrix using the category-related weights of the local features; The similarity score between two sample images is calculated based on the weighted dense connection similarity score matrix. The backbone network is then optimized based on the similarity score, and the optimized backbone network is used for few-shot image recognition. The process of using a feature propagation mechanism to convert class-independent features in the local features into class-related features, and quantifying the class relevance of the converted local features to obtain the class relevance weights of the local features, is as follows: Based on the adjacency matrix, the Laplacian matrix utilizes a feature propagation mechanism to enable each local feature to absorb complementary information from its neighbors, transforming class-independent features in the local features into class-related features: X′=PX Where P is the propagation matrix and P = (I - αL) -1 α is a balance factor and α∈(0,1); The category-related weights of the local features are defined as follows: w=X′c T Where X′ is the category-related feature, and c is the class center of the local feature.
2. The few-sample image recognition method according to claim 1, characterized in that, Before extracting local features of the sample image using the backbone network, the method further includes: The backbone network is pre-trained using mask images of the sample images for modeling.
3. The few-sample image recognition method according to claim 2, characterized in that, After extracting local features of the sample image using the backbone network, the method further includes: The extracted local features are treated as graph nodes. An adjacency matrix is constructed using these graph nodes, and the Laplacian matrix of the adjacency matrix is then built. The adjacency matrix is as follows: Where m and n represent the m-th local feature and the n-th local feature, respectively, σ represents the variance of the local feature magnitude, exp represents the exponential function, and points that are farther away are defined as neighboring points; The Laplace matrix of the adjacency matrix is: Where D is a diagonal matrix.
4. The few-sample image recognition method according to claim 3, characterized in that, The construction of the densely connected similarity score matrix is specifically as follows: Given the class-related features X1′ and X2′ of any two sample images, the dense connection similarity score matrix between X1′ and X2′ is: S[m,n]=ψ(X1′[m,:],X2′[n,:]) Where ψ(·,·) is the cosine function; The specific steps for weighting the densely connected similarity score matrix using category-related weights based on local features are as follows: S′=W1⊙S⊙W2 Where W1 is the category-related weight matrix of the local features of the first sample image, and W2 is the category-related weight matrix of the local features of the second sample image, and W2[m,:] = w2, where ⊙ represents the multiplication of elements at corresponding positions in the matrix.
5. The few-sample image recognition method according to claim 4, characterized in that, The specific steps for calculating the similarity score between two sample images based on the weighted dense connection similarity score matrix are as follows:
6. The few-sample image recognition method according to any one of claims 1 to 5, characterized in that, The specific steps for using the optimized backbone network for few-shot image recognition are as follows: The optimized backbone network is used to obtain the similarity score between the test image and any labeled image, and the label of the labeled image with the highest similarity score is assigned to the test image to generate the recognition result of the test image.
7. A few-shot image recognition apparatus utilizing the few-shot image recognition method of claim 1, characterized in that, include: Feature extraction module: used to extract local features of sample images using the backbone network; Feature transformation module: used to convert class-independent features in the local features into class-related features using a feature propagation mechanism, and to quantify the class relevance of the transformed local features to obtain the class relevance weights of the local features; Similarity score weighting module: used to construct a dense connection similarity score matrix and to weight the dense connection similarity score matrix using the category correlation weights of the local features; Model optimization module: used to calculate the similarity score between two sample images based on the weighted dense connection similarity score matrix, optimize the backbone network based on the similarity score, and use the optimized backbone network to perform few-sample image recognition.
8. A computer device, characterized in that, The computer device includes a processor and a memory coupled to the processor, wherein, The memory stores program instructions for implementing the few-sample image recognition method according to any one of claims 1-6; The processor is used to execute the program instructions stored in the memory to control the few-sample image recognition method.
9. A storage medium, characterized in that, The device stores processor-executable program instructions for performing the few-sample image recognition method according to any one of claims 1 to 6.