Image classification method and system based on local key knowledge self-distillation

Abstract

The invention discloses a self-distillation image classification method and system based on local key knowledge driving. The method comprises the steps of constructing an original image data set for an image classification task, performing standard preprocessing on original image data to enable an image to meet an input requirement of a convolutional neural network, and then dividing the data set into a training set and a test set; the original image data set comprises original images and corresponding labels; constructing a self-distillation model based on local key knowledge driving and used for realizing image classification; constructing a loss function, and training the self-distillation model by using the training set to obtain a trained self-distillation model; and inputting an image to be classified into the trained self-distillation model, and outputting a classification result. According to the method, by decomposing local and global knowledge of the target class and the non-target class, the model can be promoted to learn more key information, especially in a deeper network, so that the performance of the model is effectively improved, and the performance of the model in a fine-grained classification task is further improved.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image data processing technology, and in particular, it is an image classification method and system based on self-distillation of local key knowledge. Background Technology

[0002] With the rapid development of deep learning technology in computer vision, convolutional neural networks have achieved remarkable results in tasks such as image classification and object detection. To deploy high-performance models on mobile or edge devices, "knowledge distillation" technology is widely used for model compression and acceleration. This technique transfers knowledge learned by a complex model (teacher model) to a lightweight model (student model), enabling the student model to achieve accuracy close to that of the teacher model while maintaining low computational cost. Among these techniques, "knowledge distillation," as a special form of distillation, utilizes the model's own historical information or deep features to guide shallow learning, demonstrating excellent performance in improving the model's generalization ability.

[0003] In the knowledge distillation process, the design of "soft tags" is a key factor determining the efficiency of knowledge transfer. Existing decoupled knowledge distillation (DKD) techniques typically decouple soft tags into two parts: "target class" and "non-target class". The target class mainly reflects the difficulty of sample identification, while the non-target class contains information on the similarity between classes.

[0004] However, existing self-distillation and decoupled distillation methods still have significant technical shortcomings when processing complex background images: First, target class knowledge extraction is severely affected by background noise. Existing soft labels for target classes are usually generated based on the global prediction distribution of the entire image, lacking the ability to accurately locate key local regions (i.e., foreground targets) in the image. When the input image has a complex background (such as birds in a forest or specific people in a crowd), the model often mistakenly identifies background noise as discriminative features, resulting in the distilled "knowledge" containing a large amount of noise and reducing the accuracy of image classification. Second, non-target class smoothing lacks semantic association. Existing label smoothing methods (such as smoothing based on Zipf's law) mainly adjust the probability distribution structure from a statistical perspective, without fully considering the semantic relevance of image features. This mechanical smoothing may cause the model to be unable to distinguish between "visually similar categories" and "completely unrelated categories" when learning inter-class relationships, limiting the model's performance in fine-grained classification tasks.

[0005] Therefore, designing an image classification distillation method that can accurately extract local key features while suppressing background noise interference is a problem that urgently needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to address the problems existing in the prior art by providing a self-distillation method based on local key knowledge (LKK-SD). Through asymmetric decoupling and localized purification of knowledge, this method significantly improves model performance, thereby enhancing the model's performance in fine-grained classification tasks.

[0007] The technical solution to achieve the objective of this invention is as follows: On one hand, a self-distillation image classification method based on local key knowledge is provided, the method comprising the following steps:

[0008] Step 1: Construct the original image dataset for the image classification task, and perform standard preprocessing on the original image data to make the images meet the input requirements of the convolutional neural network. Then, divide the dataset into a training set and a test set. The original image dataset includes the original images and their corresponding labels.

[0009] Step 2: Construct a self-distillation model for image classification based on local key knowledge.

[0010] Step 3: Construct the loss function and train the self-distillation model using the training set to obtain the trained self-distillation model;

[0011] Step 4: For the image to be classified, input it into the trained self-distillation model and output the classification result.

[0012] Furthermore, the self-distillation model for image classification based on local key knowledge in step 2 includes: an image preprocessing module, an asymmetric knowledge decoupling module, a target class key feature extraction module, and a non-target class relation structure smoothing module; wherein the target class key feature extraction module and the non-target class relation structure smoothing module are executed in parallel and synchronously.

[0013] The image preprocessing module is used to preprocess the input image to generate a denoised image;

[0014] The asymmetric knowledge decoupling module is used to decouple the denoised image, divide it into target class and non-target class, and realize global knowledge extraction.

[0015] The target class key feature extraction module is used to extract local key knowledge of the target class through a pre-trained deep neural network;

[0016] The non-target class relation structure smoothing module is used to smooth the predicted probability distribution of non-target classes using Zipf's law.

[0017] Furthermore, the image preprocessing module is specifically implemented by including:

[0018] (1) For the input image, a heatmap is generated using class activation mapping technology. ;

[0019] (2) Based on the heatmap, the target region and non-target region are divided, that is, the background noise region in the image is identified and masked, specifically:

[0020] Set a response threshold ;

[0021] Will The region was identified as a background region completely unrelated to the target category. That is, the non-target area; among which, For heat map Middle position Pixel value at;

[0022] (3) For the identified background area The pixel values ​​within the image are forced to zero or replaced with the global mean to generate a denoised image. .

[0023] Furthermore, the specific implementation process of the asymmetric knowledge decoupling module includes:

[0024] (1) Denoising the image Inputting the data into a convolutional neural network yields the Logit vector for the output layer. ;

[0025] (2) Based on the output layer Logit vector Separate the tags into global target class soft tags. and global non-target class soft tags These two parts:

[0026]

[0027]

[0028] In the formula, For the first Logit vector of class value, The total number of categories, Index for the target class.

[0029] Furthermore, the training process of the deep neural network is as follows:

[0030] (1) For each sample in the training set, the target region feature map F is obtained by forward propagation through the teacher model;

[0031] (2) Divide the target region feature map F into Matrix grid;

[0032] (3) Randomly select a sub-block of the matrix grid for occlusion to form an occlusion matrix. Each element in this matrix is:

[0033]

[0034] In the formula, Represents the occlusion matrix The element in the i-th row and j-th column;

[0035] (4) The masking matrix After upsampling to the same spatial size as the feature map F of the target region, the occlusion mask is obtained. H and W represent the width and height of the feature map F of the target region, respectively.

[0036] (5) Obtain the occlusion feature map based on the occlusion mask. :

[0037]

[0038] In the formula, This indicates element-wise multiplication;

[0039] (6) Use the LIME algorithm to process the occlusion feature map Calculations are performed to obtain local key features. :

[0040]

[0041] In the formula, These are features of the i-th layer in a deep neural network. This indicates the calculation of similarity. This represents a perturbation randomly added to a selected sample;

[0042] (7) Based on local key features Generate target class soft tags :

[0043]

[0044] In the formula, express function;

[0045] (8) Based on the target class soft tag Generate target class probability That is, the target class local soft tag :

[0046]

[0047] In the formula, Represents the total number of characteristics. This represents the soft label of the target class corresponding to the feature of the j-th layer.

[0048] Furthermore, the specific implementation process of the non-target class relation structure smoothing module includes:

[0049] (1) Extract the spatial classification information of the non-target class feature map F, and map it to local classification information through a fully connected layer. And sort them in descending order; among them, The first feature map F represents the non-target class feature map. The classification information for each location is calculated using the following formula:

[0050]

[0051] In the formula, In the non-target class feature map F, the first... Feature vectors at each position Indicates a fully connected layer. express A function used to transform the output into a probability distribution;

[0052] (2) Introduce a probabilistic reconstruction operator based on Zipf's law to perform Zipf label smoothing and generate non-target class local soft labels with structured inter-class relationships. .

[0053] Furthermore, the loss function in step 3 is:

[0054]

[0055] In the formula, Indicates the total loss. This represents the consistency loss of the target class. Represents the structured loss for non-target classes. This is a hyperparameter used to balance the consistency loss of the target class and the structured loss of the non-target class.

[0056] Furthermore, the target class consistency loss for:

[0057]

[0058] In the formula, Denotes KL divergence, Indicates a global target class soft tag Local soft tags for target class The KL divergence between them.

[0059] Furthermore, the non-target class structured loss for:

[0060]

[0061] In the formula, Denotes KL divergence, Indicates a global non-target class soft tag Non-target class local soft tags The KL divergence between them.

[0062] On the other hand, an image classification system is provided, the system comprising:

[0063] The first module is used to construct the original image dataset for the image classification task, and to perform standard preprocessing on the original image data to make the images meet the input requirements of the convolutional neural network. Then, the dataset is divided into a training set and a test set. The original image dataset includes the original images and their corresponding labels.

[0064] The second module is used to construct a self-distillation model for image classification based on local key knowledge.

[0065] The third module is used to construct the loss function and train the self-distillation model using the training set to obtain the trained self-distillation model.

[0066] The fourth module is used to input the image to be classified into the trained self-distillation model and output the classification result.

[0067] Compared with the prior art, the significant advantages of this invention are:

[0068] (1) Improved model performance: This invention effectively improves the classification accuracy of student models on various network architectures and benchmark datasets by accurately extracting key knowledge, which is superior to many existing advanced self-distillation methods.

[0069] (2) Knowledge-oriented learning: By background denoising and target class knowledge localization (LIME combined with random masking), this method can force the model to focus on the most discriminative local features in the image, thereby improving the purity and effectiveness of knowledge transfer.

[0070] (3) Optimize inter-class relationships: Based on Zipf's law, the label smoothing process provides a structured and more reasonable distribution guidance for the transfer of non-target class knowledge, avoiding the randomness of knowledge transfer.

[0071] (4) Each step (background denoising, knowledge decoupling, asymmetric localization, and double loss) has been verified by ablation experiments and works together to ensure the robustness and high performance of the overall framework of the present invention, thereby improving the accuracy of image classification.

[0072] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0073] Figure 1 This is a flowchart of the self-distillation process based on local criticality knowledge in one embodiment of the present invention.

[0074] Figure 2 This is a schematic diagram of the overall principle architecture of a self-distillation method based on local key knowledge (LKK-SD) in one embodiment.

[0075] Figure 3 This is a schematic diagram illustrating background area masking using a CAM heatmap in one embodiment.

[0076] Figure 4 This is a schematic diagram of one embodiment where soft tags are decoupled into global target class and global non-target class soft tags.

[0077] Figure 5 This is a schematic diagram of a target class local key knowledge extraction framework in one embodiment.

[0078] Figure 6 This is a schematic diagram of the framework for extracting local key knowledge of non-target classes (Zipf label smoothing) in one embodiment. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0080] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0081] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0082] In one embodiment, combined Figure 1 This paper provides a self-distillation image classification method based on local key knowledge, the method comprising the following steps:

[0083] Step 1: Construct the original image dataset for the image classification task, and perform standard preprocessing (size normalization, data augmentation, etc.) on the original image data to make the images meet the input requirements of the convolutional neural network. Then, divide the dataset into training set and test set; the original image dataset includes the original images and their corresponding labels.

[0084] Step 2: Construct a self-distillation model for image classification based on local key knowledge.

[0085] Step 3: Construct the loss function and train the self-distillation model using the training set to obtain the trained self-distillation model;

[0086] Step 4: For the image to be classified, input it into the trained self-distillation model and output the classification result.

[0087] Furthermore, combined Figures 2 to 6 The self-distillation model for image classification based on local key knowledge in step 2 includes: an image preprocessing module, an asymmetric knowledge decoupling module, a target class key feature extraction module, and a non-target class relation structure smoothing module; wherein the target class key feature extraction module and the non-target class relation structure smoothing module are executed in parallel and synchronously.

[0088] The image preprocessing module is used to preprocess the input image to generate a denoised image; here, this module is used to physically shield low-response background regions before the image is input into the network to prevent the model from relying on false background features.

[0089] The asymmetric knowledge decoupling module is used to decouple the denoised image, divide it into target class and non-target class, and realize global knowledge extraction.

[0090] The target class key feature extraction module is used to extract local key knowledge of the target class through a pre-trained deep neural network;

[0091] The non-target class relation structure smoothing module is used to smooth the predicted probability distribution of non-target classes using Zipf's law.

[0092] Preferably, in some embodiments, the image preprocessing module specifically includes the following steps:

[0093] (1) For the input image, a heatmap is generated using class activation mapping technology. ;

[0094] (2) Based on the heatmap, the target region and non-target region are divided, that is, the background noise region in the image is identified and masked, specifically:

[0095] Set a response threshold ;

[0096] Will The region was identified as a background region completely unrelated to the target category. That is, the non-target area; among which, For heat map Middle position Pixel value at;

[0097] (3) For the identified background area The pixel values ​​within the image are forced to zero or replaced with the global mean to generate a denoised image. .

[0098] Preferably, in some embodiments, the asymmetric knowledge decoupling module is specifically implemented by including:

[0099] (1) Denoising the image Inputting the data into a convolutional neural network yields the Logit vector for the output layer. ;

[0100] (2) Based on the output layer Logit vector Separate the tags into global target class soft tags. and global non-target class soft tags These two parts:

[0101]

[0102]

[0103] In the formula, For the first Logit vector of class value, The total number of categories, Index for the target class.

[0104] Here, the global target class soft tag It only includes the predicted probability of the target class; global non-target class soft labels. : Includes the predicted probability distribution of all categories except the target class.

[0105] Preferably, in some embodiments, the training process of the deep neural network is as follows:

[0106] (1) For each sample in the training set, the target region feature map F is obtained by forward propagation through the teacher model;

[0107] (2) Divide the target region feature map F into Matrix grid;

[0108] (3) Randomly select a sub-block of the matrix grid for occlusion to form an occlusion matrix. Each element in this matrix is:

[0109]

[0110] In the formula, Represents the occlusion matrix The element in the i-th row and j-th column;

[0111] (6) The masking matrix After upsampling to the same spatial size as the feature map F of the target region, the occlusion mask is obtained. H and W represent the width and height of the feature map F of the target region, respectively.

[0112] (7) Obtain the occlusion feature map based on the occlusion mask. :

[0113]

[0114] In the formula, This indicates element-wise multiplication;

[0115] (6) Use the LIME algorithm to process the occlusion feature map Calculations are performed to obtain local key features. (Simulating a scenario of local feature loss):

[0116]

[0117] In the formula, These are features of the i-th layer in a deep neural network. This indicates the calculation of similarity. This represents a perturbation randomly added to a selected sample;

[0118] (7) Based on local key features Generate target class soft tags :

[0119]

[0120] In the formula, express function;

[0121] (8) Based on the target class soft tag Generate target class probability That is, the target class local soft tag :

[0122]

[0123] In the formula, Represents the total number of characteristics. This represents the soft label of the target class corresponding to the feature of the j-th layer.

[0124] Preferably, in some embodiments, the non-target class relationship structure smoothing module is specifically implemented by including:

[0125] (1) Extract the spatial classification information of the non-target class feature map F, and map it to local classification information through a fully connected layer. And sort them in descending order; among them, The first feature map F represents the non-target class feature map. The classification information for each location is calculated using the following formula:

[0126]

[0127] In the formula, In the non-target class feature map F, the first... Feature vectors at each position Indicates a fully connected layer. express A function used to transform the output into a probability distribution;

[0128] (2) Introduce a probabilistic reconstruction operator based on Zipf's law to perform Zipf label smoothing and generate non-target class local soft labels with structured inter-class relationships. This is to enhance the model's ability to distinguish hard negative samples (similar categories).

[0129] Preferably, in some embodiments, the loss function in step 3 is:

[0130]

[0131] In the formula, Indicates the total loss. This represents the consistency loss of the target class. Represents the structured loss for non-target classes. This is a hyperparameter used to balance the consistency loss of the target class and the structuring loss of the non-target class. Preferably, The preferred setting is 0.5.

[0132] Here, the target class consistency loss for:

[0133]

[0134] In the formula, Denotes KL divergence, Indicates a global target class soft tag Local soft tags for target class The KL divergence between them.

[0135] Here, the non-target class structured loss for:

[0136]

[0137] In the formula, Denotes KL divergence, Indicates a global non-target class soft tag Non-target class local soft tags The KL divergence between them.

[0138] Preferably, in some embodiments, step 3 involves training the self-distillation model using a training set, specifically including: inputting images into the model constructed in step 2 using the training dataset built in step 1. During forward propagation, image denoising, knowledge decoupling, random occlusion and LIME extraction, and Zipf smoothing are performed sequentially, and the total loss Lossall is calculated. The gradient is calculated using the backpropagation algorithm, and the parameters of the student model are updated until the model converges or reaches a preset number of training rounds. After training, a lightweight image classification model is obtained.

[0139] Preferably, in some embodiments, step 4 specifically includes:

[0140] Image input: Obtain one or more image data to be classified.

[0141] Inference computation: The image to be classified is input into the lightweight model trained in step 3. The model uses the trained weight parameters to extract features and perform computation on the input image.

[0142] Output results: The model outputs the probability values ​​of the image belonging to each category, and selects the category with the highest probability as the final classification result and outputs it.

[0143] In one embodiment, an image classification system is provided, the system comprising:

[0144] The first module is used to construct the original image dataset for the image classification task, and to perform standard preprocessing on the original image data to make the images meet the input requirements of the convolutional neural network. Then, the dataset is divided into a training set and a test set. The original image dataset includes the original images and their corresponding labels.

[0145] The second module is used to construct a self-distillation model for image classification based on local key knowledge.

[0146] The third module is used to construct the loss function and train the self-distillation model using the training set to obtain the trained self-distillation model.

[0147] The fourth module is used to input the image to be classified into the trained self-distillation model and output the classification result.

[0148] Specific limitations regarding the image classification system based on self-distillation of local key knowledge can be found in the limitations of the image classification method based on self-distillation of local key knowledge mentioned above, and will not be repeated here. Each module in the aforementioned image classification system based on self-distillation of local key knowledge can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0149] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements:

[0150] Step 1: Construct the original image dataset for the image classification task, and perform standard preprocessing on the original image data to make the images meet the input requirements of the convolutional neural network. Then, divide the dataset into a training set and a test set. The original image dataset includes the original images and their corresponding labels.

[0151] Step 2: Construct a self-distillation model for image classification based on local key knowledge.

[0152] Step 3: Construct the loss function and train the self-distillation model using the training set to obtain the trained self-distillation model;

[0153] Step 4: For the image to be classified, input it into the trained self-distillation model and output the classification result.

[0154] For specific limitations on each step, please refer to the limitations of the image classification method based on self-distillation of local key knowledge mentioned above, which will not be repeated here.

[0155] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being implemented when executed by a processor:

[0156] Step 1: Construct the original image dataset for the image classification task, and perform standard preprocessing on the original image data to make the images meet the input requirements of the convolutional neural network. Then, divide the dataset into a training set and a test set. The original image dataset includes the original images and their corresponding labels.

[0157] Step 2: Construct a self-distillation model for image classification based on local key knowledge.

[0158] Step 3: Construct the loss function and train the self-distillation model using the training set to obtain the trained self-distillation model;

[0159] Step 4: For the image to be classified, input it into the trained self-distillation model and output the classification result.

[0160] For specific limitations on each step, please refer to the limitations of the image classification method based on self-distillation of local key knowledge mentioned above, which will not be repeated here.

[0161] As a specific example, the invention will be further verified and illustrated in one embodiment.

[0162] Based on the existing CIFAR-100 and ImageNet datasets, the accuracy of LKK-SD (the self-distillation method driven by local key knowledge in this invention) is compared with other manually set soft-label self-distillation methods, as shown in Table 1 below:

[0163] Table 1 Comparison of Experimental Results

[0164]

[0165] As shown in Table 1, our proposed method achieves state-of-the-art performance on the ResNet50 model on ImageNet, exceeding the baseline accuracy by 0.54%. Furthermore, on ResNet18 for CIFAR-100, ResNet18 for ImageNet, and MobileNet, our method also achieves comparable performance to the current state-of-the-art methods, with significant improvements of 1.24%, 0.84%, and 1.11% respectively compared to the baseline. These results strongly demonstrate that decomposing the local and global knowledge of target and non-target classes enables the model to learn more crucial information, especially in deeper network layers, thereby effectively improving its performance and ultimately enhancing its performance in fine-grained classification tasks.

[0166] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention without departing from its spirit and scope should be included within the protection scope of the present invention.

Claims

1. A self-distillation image classification method based on local key knowledge, characterized in that, The method includes the following steps: Step 1: Construct the original image dataset for the image classification task, and perform standard preprocessing on the original image data to make the images meet the input requirements of the convolutional neural network. Then, divide the dataset into a training set and a test set. The original image dataset includes the original images and their corresponding labels. Step 2: Construct a self-distillation model for image classification based on local key knowledge. Step 3: Construct the loss function and train the self-distillation model using the training set to obtain the trained self-distillation model; Step 4: For the image to be classified, input it into the trained self-distillation model and output the classification result.

2. The image classification method based on self-distillation of local key knowledge according to claim 1, characterized in that, The self-distillation model for image classification based on local key knowledge described in step 2 includes: an image preprocessing module, an asymmetric knowledge decoupling module, a target class key feature extraction module, and a non-target class relation structure smoothing module; wherein the target class key feature extraction module and the non-target class relation structure smoothing module are executed in parallel and synchronously. The image preprocessing module is used to preprocess the input image to generate a denoised image; The asymmetric knowledge decoupling module is used to decouple the denoised image, divide it into target class and non-target class, and realize global knowledge extraction. The target class key feature extraction module is used to extract local key knowledge of the target class through a pre-trained deep neural network; The non-target class relation structure smoothing module is used to smooth the predicted probability distribution of non-target classes using Zipf's law.

3. The image classification method based on self-distillation of local key knowledge according to claim 2, characterized in that, The image preprocessing module is specifically implemented as follows: (1) For the input image, a heatmap is generated using class activation mapping technology. ; (2) Based on the heatmap, the target area and non-target area are divided, that is, the background noise area in the image is identified and masked, specifically: Set a response threshold ; Will The region was identified as a background region completely unrelated to the target category. That is, the non-target area; among which, For heat map Middle position Pixel value at; (3) For the identified background area The pixel values ​​within the image are forced to zero or replaced with the global mean to generate a denoised image. .

4. The image classification method based on self-distillation of local key knowledge according to claim 3, characterized in that, The specific implementation process of the asymmetric knowledge decoupling module includes: (1) Denoising the image Inputting the data into a convolutional neural network yields the Logit vector for the output layer. ; (2) Based on the output layer Logit vector Separate the tags into global target class soft tags. and global non-target class soft tags These two parts: ； ； In the formula, For the first Logit vector of class value, The total number of categories, Index for the target class.

5. The image classification method based on self-distillation of local key knowledge according to claim 2, characterized in that, The training process of the deep neural network is as follows: (1) For each sample in the training set, the target region feature map F is obtained by forward propagation through the teacher model; (2) Divide the target region feature map F into Matrix grid; (3) Randomly select a sub-block of the matrix grid for occlusion to form an occlusion matrix. Each element in this matrix is: ； In the formula, Represents the occlusion matrix The element in the i-th row and j-th column; (4) The masking matrix After upsampling to the same spatial size as the feature map F of the target region, the occlusion mask is obtained. ; H and W are the width and height of the feature map F of the target region, respectively; (5) Obtain the occlusion feature map based on the occlusion mask. : ； In the formula, This indicates element-wise multiplication; (6) Use the LIME algorithm to process the occlusion feature map Calculations are performed to obtain local key features. : ； In the formula, These are features of the i-th layer in a deep neural network. This indicates the calculation of similarity. This represents a perturbation randomly added to a selected sample; (7) Based on local key features Generate target class soft tags : ； In the formula, express function; (8) Based on the target class soft tag Generate target class probability That is, the target class local soft tag : ； In the formula, Represents the total number of characteristics. This represents the soft label of the target class corresponding to the feature of the j-th layer.

6. The image classification method based on self-distillation of local key knowledge according to claim 5, characterized in that, The implementation process of the non-target class relation structure smoothing module includes: (1) Extract the spatial classification information of the non-target class feature map F, and map it to local classification information through a fully connected layer. And sort them in descending order; among them, The first feature map F represents the non-target class feature map. The classification information for each location is calculated using the following formula: ； In the formula, In the non-target class feature map F, the first... Feature vectors at each position Indicates a fully connected layer. express A function used to transform the output into a probability distribution; (2) Introduce a probabilistic reconstruction operator based on Zipf's law to perform Zipf label smoothing and generate non-target class local soft labels with structured inter-class relationships. .

7. The image classification method based on self-distillation of local key knowledge according to claim 6, characterized in that, The loss function in step 3 is: ； In the formula, Indicates the total loss. This represents the consistency loss of the target class. Represents the structured loss for non-target classes. This is a hyperparameter used to balance the consistency loss of the target class and the structured loss of the non-target class.

8. The image classification method based on self-distillation of local key knowledge according to claim 7, characterized in that, The target class consistency loss for: ； In the formula, Indicates the KL divergence. Indicates a global target class soft tag Local soft tags for target class The KL divergence between them.

9. The image classification method based on self-distillation of local key knowledge according to claim 7, characterized in that, The non-target class structured loss for: ； In the formula, Indicates the KL divergence. Indicates a global non-target class soft tag Non-target class local soft tags The KL divergence between them.

10. An image classification system based on the method of any one of claims 1 to 9, characterized in that, The system includes: The first module is used to construct the original image dataset for the image classification task, and to perform standard preprocessing on the original image data to make the images meet the input requirements of the convolutional neural network. Then, the dataset is divided into a training set and a test set. The original image dataset includes the original images and their corresponding labels. The second module is used to construct a self-distillation model for image classification based on local key knowledge. The third module is used to construct the loss function and train the self-distillation model using the training set to obtain the trained self-distillation model. The fourth module is used to input the image to be classified into the trained self-distillation model and output the classification result.

Images