A method and system for reusing convolutional discarded information for fine-grained image classification with small samples

Abstract

This invention discloses a method and system for reusing discarded convolutional information in small-sample fine-grained image classification, belonging to the field of computer vision. The method includes the following steps: acquiring images of a support set and a query set, and extracting the backbone feature descriptors of the images using a convolutional neural network; after at least one convolutional block in the convolutional neural network, retrieving the discarded feature information of the convolutional block through a feature retrieval path to generate a retrieved feature descriptor; fusing the backbone feature descriptor with the retrieved feature descriptor to obtain a high-dimensional feature descriptor in a unified format; calculating the similarity between the images of the query set and the support set based on the high-dimensional feature descriptor, and performing small-sample fine-grained image classification based on the similarity, thereby realizing the reuse of discarded convolutional information. This invention can effectively retrieve discarded feature information in convolutional neural networks, improving the performance of small-sample fine-grained image classification.

Description

Technical Field

[0001] This invention belongs to the field of computer vision, specifically relating to a method and system for reusing convolutional discarded information for small-sample fine-grained image classification. Background Technology

[0002] Fine-grained image classification aims to distinguish subclasses belonging to the same basic category with fine precision, and has significant application value in fields such as biometrics and industrial inspection. However, acquiring large amounts of high-quality labeled data is costly. Therefore, few-shot fine-grained image classification (FSFGIC) has emerged, its core being the use of a very small number of labeled samples to learn feature representations that can distinguish subtle differences between classes. Currently, mainstream methods in this field typically employ pre-trained convolutional neural network (CNN) models (such as AlexNet, Conv-4, ResNet-12, etc.) as the backbone network for feature extraction. These deep convolutional network-based methods can automatically learn hierarchical features and have become an effective technical approach for solving few-shot fine-grained classification tasks.

[0003] However, the aforementioned existing technologies have inherent limitations in practical implementation. First, due to the structural characteristics of convolutional neural networks, some feature information is inevitably lost during feature abstraction and dimensionality reduction through operations such as convolution and pooling. These discarded features may contain detailed information valuable for fine-grained category discrimination. Existing methods generally do not consider how to recover and utilize this discarded feature information, resulting in limited feature representation capabilities. Second, most FSFGIC methods only utilize the mainstream features output by CNNs, without paying attention to the potential information discarded due to network structure, parameter settings, or intermediate layer operations. Ignoring this information may lead to insufficient ability to discriminate subtle differences between categories, limiting the model's classification performance in small sample scenarios.

[0004] Due to the aforementioned shortcomings, the existing FSFGIC method still has significant room for improvement in classification accuracy and generalization ability in real-world applications with very few samples and fine class distinctions. Therefore, effectively recovering and fully utilizing the feature information discarded in CNNs has become one of the key issues in improving the performance of fine-grained image classification with few samples. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for reusing discarded information from convolutional neural networks for small-sample fine-grained image classification, which can effectively recover discarded feature information in convolutional neural networks and improve the performance of small-sample fine-grained image classification.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, a method for reusing convolutional discarded information for small-sample fine-grained image classification includes the following steps: Acquire images from the support set and query set, and extract the backbone feature descriptors of the images using a convolutional neural network; After at least one convolutional block in the convolutional neural network, the discarded feature information of the convolutional block is recycled through a feature recycling path to generate recycled feature descriptors. The backbone feature descriptor is fused with the recycled feature descriptor to obtain a high-dimensional feature descriptor in a unified format; Based on the high-dimensional feature descriptor, the similarity between the images in the query set and the support set is calculated, and small-sample fine-grained image classification is performed according to the similarity, realizing the reuse of convolution discarded information.

[0007] In some implementations, the step of recovering the discarded feature information of the convolutional block through a feature recovery path specifically includes: The discarded feature information is transformed and enhanced by sequentially passing through convolutional layers, pixel unshuffle operations, and batch normalization layers to obtain recycled feature descriptors.

[0008] In some implementations, the recycling process employs multi-round recursive recycling, specifically including: in each round of recycling, the discarded feature information obtained from the previous round of recycling is input again into the feature recycling path for recycling.

[0009] In some implementations, the step of fusing the backbone feature descriptor with the recycling feature descriptor specifically includes: The backbone feature descriptor and the recycled feature descriptor are concatenated along the channel dimension, and then channel integration is performed through convolution. Subsequently, normalization and pooling are performed to obtain a high-dimensional feature descriptor in a unified format.

[0010] In some implementations, the convolutional neural network is a Conv-4 structure or a ResNet-12 structure.

[0011] In some implementations, the step of calculating the similarity between images in the query set and the support set based on the high-dimensional feature descriptor, and performing small-sample fine-grained image classification based on the similarity, specifically includes: Based on the high-dimensional feature descriptor, the similarity between the images in the query set and the support set is calculated by Euclidean distance or dot product attention, and normalized by the softmax function to obtain the classification probability. The preset category corresponding to the largest classification probability is selected as the small sample fine-grained image classification result.

[0012] Secondly, a system for reusing convolutional discarded information for small-sample fine-grained image classification includes: An image input module is used to acquire images from a support set and a query set, and to extract the backbone feature descriptors of the images through a convolutional neural network. The feature retrieval module is used to retrieve discarded feature information of at least one convolutional block in the convolutional neural network through a feature retrieval path, and generate a retrieved feature descriptor. The feature fusion module is used to fuse the backbone feature descriptor with the recycled feature descriptor to obtain a high-dimensional feature descriptor in a unified format; The similarity calculation and classification module is used to calculate the similarity between the images in the query set and the support set based on the high-dimensional feature descriptor, and to perform fine-grained image classification for small samples based on the similarity, thereby realizing the reuse of information discarded by convolution.

[0013] Thirdly, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable in the processor, wherein the processor, when executing the computer program, implements the steps of the method for reusing convolutional discarded information for small-sample fine-grained image classification.

[0014] Fourthly, a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method for reusing convolutional discarded information for small-sample fine-grained image classification.

[0015] Fifthly, a computer program product comprising a computer program that, when executed by a processor, implements the steps of the convolutional discard information reuse method for small-sample fine-grained image classification.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention actively retrieves and processes the feature information output by convolutional blocks through a feature retrieval pathway, and fuses it with the backbone features. This solves the problem of insufficient utilization of fine-grained information caused by feature discarding in the prior art. It enables the model to capture and retain detailed features that are crucial for class differentiation from the intermediate layers of the network, thereby significantly enhancing the discriminative power and expressive richness of features under conditions of scarce samples, and improving the overall performance of fine-grained classification with small samples.

[0017] Furthermore, the discarded feature information is transformed and enhanced sequentially through convolutional layers, pixel unshuffle operations, and batch normalization layers to obtain recovered feature descriptors. The convolutional layers efficiently integrate cross-channel information and achieve dimensionality reduction, the pixel unshuffle operation actively restores spatial detail structure, effectively compensating for information loss caused by operations such as pooling, and batch normalization stabilizes the distribution of the recovered discarded feature information. Together, they ensure that the recovered feature descriptors are extracted and transmitted with high quality and stability.

[0018] Furthermore, in each round of recycling, the discarded feature information obtained in the previous round of recycling is input again into the feature recycling path for recycling, which greatly enriches the hierarchy and information capacity of the final feature representation. Especially in small sample scenarios, it can make fuller use of limited data and further alleviate the information bottleneck problem. Attached Figure Description

[0019] Figure 1 A flowchart of a method for reusing convolutional discarded information for small-sample fine-grained image classification provided by an embodiment of the present invention; Figure 2 This is a structural diagram of a convolutional discard information reuse system for small-sample fine-grained image classification provided in an embodiment of the present invention; Figure 3 A schematic diagram of the FIR-Net structure based on the Conv-4 backbone provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the FIR-Net structure based on the ResNet-12 backbone provided in an embodiment of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. The content described herein is for explanation rather than limitation of the present invention.

[0021] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification of this invention are intended to cover a non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, systems, products or devices.

[0022] like Figure 1 As shown, this embodiment provides a method for reusing convolutional discarded information for small-sample fine-grained image classification, including the following steps: S1, acquire images of the support set and query set, and extract the backbone feature descriptors of the images through a convolutional neural network; Specifically, features are extracted from each input image in the support and query sets using a convolutional neural network. Conv-4 or ResNet-12 can be used as the backbone network. Conv-4 is a shallow convolutional neural network consisting of four convolutional blocks, each containing a 3×3 convolutional layer, a batch normalization layer, a Leaky ReLU activation function, and a 2×2 max-pooling layer. ResNet-12 is a residual network structure containing twelve layers. After the above processing, preliminary backbone feature descriptors can be obtained.

[0023] S2, after at least one convolutional block in the convolutional neural network, the discarded feature information of the convolutional block is recycled through a feature recycling path to generate a recycled feature descriptor; A feature retrieval path is set after at least one convolutional block of the convolutional neural network; in some embodiments, feature retrieval paths can be set after multiple key convolutional blocks to retrieve features that were not output by the backbone network during the convolutional operation and would otherwise be discarded. This feature retrieval path includes 1×1 convolutions (i.e., convolution operations with a kernel size of 1×1), pixel unshuffle operations (i.e., inverse pixel rearrangement, restoring the feature map to a finer-grained spatial structure), and batch normalization layer processing. Each retrieval path can extract a set of retrieved feature descriptors (i.e., feature vectors used for subsequent fusion). The retrieval process can employ multiple rounds of recursive retrieval; in each round, the features obtained from the previous round are input back into the feature retrieval path for processing to extract richer, more detailed features.

[0024] S3, the backbone feature descriptor and the recycled feature descriptor are fused to obtain a high-dimensional feature descriptor in a unified format; The feature descriptors output from the backbone network are concatenated with the feature descriptors retrieved from each level in a predetermined order (i.e., all features are merged along the channel dimension). The concatenated features are then integrated through a 1×1 convolution operation, followed by normalization and pooling, ultimately yielding a high-dimensional feature descriptor in a unified format. This fused feature incorporates both the mainstream information from the backbone network and the fine-grained discriminative information obtained through the feature retrieval pathway.

[0025] S4. Based on the high-dimensional feature descriptor, calculate the similarity between the images in the query set and the support set, and perform small-sample fine-grained image classification according to the similarity to realize the reuse of convolution discarded information.

[0026] The final feature descriptors of all support sets and query sets are input into the similarity calculation module. Euclidean distance, dot product attention, and other methods are used to calculate the distance between the query sample and each category of support set. The formula is as follows:

[0027] in, Indicates query sample With supporting samples The Euclidean distance between feature descriptors.

[0028] The distance scores for all categories are normalized using the softmax function (i.e., the normalized exponential function) to obtain the final classification probability, as shown in the following formula:

[0029] in, Indicates query sample Belongs to the Classification probability of a class Indicates the first Feature descriptors of class support set samples Indicates the total number of categories. Indicates query sample With the The distance between class support samples is used. Since a smaller distance indicates a higher similarity, the negative value of the distance is used for normalization in softmax to obtain the final classification probability of each class.

[0030] This method can be trained end-to-end (i.e., input the original image, output the final classification result, and optimize all parameters jointly), and is applicable to various publicly available fine-grained image datasets such as CUB-200-2011 (Caltech Bird Image Dataset) and Stanford Dogs (Stanford University Dog Image Dataset), achieving efficient small-sample fine-grained image classification.

[0031] Figure 3 The structure and data flow of the Feature Information Reharvesting Network (FIR-Net) based on the Conv-4 backbone are demonstrated. Figure 3 (a) shows the Conv-4 backbone structure, which consists of four concatenated convolutional blocks, each of which includes a 3×3 convolutional layer, a batch normalization layer, a Leaky ReLU activation layer and a 2×2 max pooling layer in sequence. Figure 3 (b) illustrates an exemplary overall flow of FIR-Net based on the Conv-4 backbone in a 5-way 1-shot task. Support set on the left. and query images Each input sample is input into a feature extraction network with shared parameters, and each input sample is uniformly denoted as... Main roads and secondary roads First, the input Perform standard convolutional feature extraction to obtain the backbone feature descriptor. To recover information weakened or discarded during early convolutional filtering, this implementation sets up multiple information recovery branches in addition to the main branch. Specifically, the initial information tensor is generated from the input features through 1×1 convolution and batch normalization. and intermediate filter features and Perform differential operations step by step to obtain the subsequent information tensor. and The recycling relationship can be expressed as:

[0032] Meanwhile, another retrieval path generates an initial information tensor through spatial scale alignment, 1×1 convolution, and batch normalization. and intermediate filter features and Perform differential operations step by step to obtain the subsequent information tensor. and The recycling relationship can be expressed as:

[0033] in, This indicates the corresponding recycling path or stage index. This represents a tensor containing information to be reclaimed in the current reclamation path. This represents the intermediate filtered features extracted from the corresponding layer of the main branch. The difference between the two represents the fine-grained information remaining after the current convolutional filtering, which can still be reused later. The resulting information tensors are then fed into the recycling branches. , , and The corresponding recycling feature descriptors are extracted. , , and Furthermore, the intermediate information tensors in the recycling path can undergo spatial-to-channel information rearrangement via a pixel unshuffle operation before entering subsequent recycling branches, preserving the fine-grained local structural information remaining after convolutional filtering. The feature descriptors output from each branch are finally concatenated and fused along the channel dimension to form a unified high-dimensional feature descriptor. Then, query the sample features. Supporting sample features to Input the Similarity Measure module on the right to calculate similarity, output the corresponding classification scores, and use the category with the highest score as the query image. The classification results are shown in the figure. This figure fully illustrates the entire process from backbone feature extraction, information discarding and recovery, to feature fusion and similarity classification under the shallow Conv-4 backbone.

[0034] Figure 4 This paper demonstrates the structure and feature retrieval process of a Feature Information Retrieval Network (FIR-Net) based on the ResNet-12 backbone. Figure 4 (a) shows the ResNet-12 backbone structure, which consists of four cascaded residual stages. Each residual stage includes multiple convolutional units and a residual bypass branch. The convolutional units sequentially include a 3×3 convolutional layer, a batch normalization layer, a Leaky ReLU activation layer, and a 2×2 max pooling layer. The “4×” in the figure indicates that this type of residual structure is stacked repeatedly in stages. Figure 4 (b) shows the FIR structure after integrating the feature information recovery path on the ResNet-12 backbone. For any input image Main residual branch First, extract the backbone features to obtain the main feature descriptor. To recover information weakened or discarded by convolutional filtering within the residual block, an initial information tensor is generated from the input features. And in different levels of the main residual branch and intermediate filtering characteristics , , and The retrieval relationship of performing differential operations at each level can be expressed as:

[0035] in, This indicates the corresponding recycling path or stage index. Indicates the first The tensor of information to be recycled input in each recycling stage. This represents the difference between the intermediate filter features extracted from the corresponding layer of the main residual branch and the two features. This represents the remaining information to be recycled after filtering in this layer, and it continues to be input into subsequent recycling branches. This process yields the subsequent information tensors in sequence. , , and They were then sent to the recycling branch lines respectively. , , and The corresponding recycling feature descriptors are extracted. , , and The features output from the main residual branch and each recovery branch are finally concatenated and fused along the channel dimension, and then integrated through subsequent convolutions to obtain a unified final feature descriptor. Compared to the conventional ResNet-12 approach that only utilizes residual short-circuit branches for feature transfer, this implementation explicitly separates, recovers, and reuses fine-grained discriminative information weakened or discarded by convolutional filtering, beyond the residual connections. This allows the semantic information retained by the main path and the detailed information supplemented by the recovered path to jointly participate in feature representation construction, thereby improving the performance of small-sample fine-grained image classification and demonstrating that the proposed solution can be adapted to implementations of deeper residual backbone networks.

[0036] like Figure 2 As shown, this embodiment provides a system for reusing convolutional discarded information for small-sample fine-grained image classification, including: An image input module is used to acquire images from a support set and a query set, and to extract the backbone feature descriptors of the images through a convolutional neural network. The feature retrieval module is used to retrieve discarded feature information of at least one convolutional block in the convolutional neural network through a feature retrieval path, and generate a retrieved feature descriptor. The feature fusion module is used to fuse the backbone feature descriptor with the recycled feature descriptor to obtain a high-dimensional feature descriptor in a unified format; The similarity calculation and classification module is used to calculate the similarity between the images in the query set and the support set based on the high-dimensional feature descriptor, and to perform fine-grained image classification for small samples based on the similarity, thereby realizing the reuse of information discarded by convolution.

[0037] The module division in this embodiment of the invention is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in the various embodiments of the invention can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0038] This embodiment also provides a computer device, which includes a processor and a memory. The memory is used to store a computer program (in this embodiment, the computer program includes computational components and iterative components, capable of model calculation and model updating). The computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the computer storage medium to realize the corresponding method flow or corresponding function. The processor described in this embodiment can be used to operate a convolution discard information reuse method for small-sample fine-grained image classification.

[0039] This embodiment also provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the convolution discard information reuse method for small-sample fine-grained image classification in the above embodiment.

[0040] This embodiment also provides a computer program product, which includes a computer program that, when executed by a processor, implements the corresponding steps of the convolution discard information reuse method for small-sample fine-grained image classification described in the above embodiment.

[0041] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0042] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0043] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0044] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0045] 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 it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for reusing convolutional discarded information in fine-grained image classification with small samples, characterized in that, Includes the following steps: Acquire images from the support set and query set, and extract the backbone feature descriptors of the images using a convolutional neural network; After at least one convolutional block in the convolutional neural network, the discarded feature information of the convolutional block is recycled through a feature recycling path to generate recycled feature descriptors. The backbone feature descriptor is fused with the recycled feature descriptor to obtain a high-dimensional feature descriptor in a unified format; Based on the high-dimensional feature descriptor, the similarity between the images in the query set and the support set is calculated, and small-sample fine-grained image classification is performed according to the similarity, realizing the reuse of convolution discarded information.

2. The method for reusing convolutional discarded information for small-sample fine-grained image classification according to claim 1, characterized in that, The step of recovering the discarded feature information of the convolutional block through the feature recovery path specifically includes: The discarded feature information is transformed and enhanced by sequentially passing through convolutional layers, pixel unshuffle operations, and batch normalization layers to obtain recycled feature descriptors.

3. The method for reusing convolutional discarded information for small-sample fine-grained image classification according to claim 2, characterized in that, The recycling process employs a multi-round recursive recycling method, specifically including: in each round of recycling, the discarded feature information obtained from the previous round of recycling is input again into the feature recycling path for recycling.

4. The method for reusing convolutional discarded information for small-sample fine-grained image classification according to claim 1, characterized in that, The step of fusing the backbone feature descriptor with the recycled feature descriptor specifically includes: The backbone feature descriptor and the recycled feature descriptor are concatenated along the channel dimension, and then channel integration is performed through convolution. Subsequently, normalization and pooling are performed to obtain a high-dimensional feature descriptor in a unified format.

5. The method for reusing convolutional discarded information for small-sample fine-grained image classification according to claim 1, characterized in that, The convolutional neural network is either a Conv-4 structure or a ResNet-12 structure.

6. The method for reusing convolutional discarded information for small-sample fine-grained image classification according to claim 1, characterized in that, The step of calculating the similarity between images in the query set and the support set based on the high-dimensional feature descriptor, and performing fine-grained image classification for small samples based on the similarity, specifically includes: Based on the high-dimensional feature descriptor, the similarity between the images in the query set and the support set is calculated by Euclidean distance or dot product attention, and normalized by the softmax function to obtain the classification probability. The preset category corresponding to the largest classification probability is selected as the small sample fine-grained image classification result.

7. A system for reusing convolutional discarded information for small-sample fine-grained image classification, characterized in that, include: An image input module is used to acquire images from a support set and a query set, and to extract the backbone feature descriptors of the images through a convolutional neural network. The feature retrieval module is used to retrieve discarded feature information of at least one convolutional block in the convolutional neural network through a feature retrieval path, and generate a retrieved feature descriptor. The feature fusion module is used to fuse the backbone feature descriptor with the recycled feature descriptor to obtain a high-dimensional feature descriptor in a unified format; The similarity calculation and classification module is used to calculate the similarity between the images in the query set and the support set based on the high-dimensional feature descriptor, and to perform fine-grained image classification for small samples based on the similarity, thereby realizing the reuse of information discarded by convolution.

8. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable in the processor, wherein the processor executes the computer program to implement the steps of the method for reusing convolutional discarded information for small-sample fine-grained image classification as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the convolution discard information reuse method for small-sample fine-grained image classification as described in any one of claims 1 to 6.

10. A computer program product, the computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the convolution discard information reuse method for small sample fine-grained image classification as described in any one of claims 1 to 6.

Images