SAR image recognition method based on self-knowledge distillation
By using self-knowledge distillation technology to generate multiple sub-models and transfer knowledge in SAR image recognition, the problem of insufficient computing resources on edge devices is solved, the recognition efficiency and accuracy are improved, and the implementation cost is reduced.
Patent Information
- Application Number
- CN202411517609.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-29
- Publication Date
- 2025-01-14
AI Technical Summary
Existing SAR image recognition methods require significant computational resources and complex model adjustments on edge devices, making it difficult to efficiently process remote sensing images.
By employing self-knowledge distillation, multiple sub-models are generated within a single network and knowledge is transferred between them. Random deactivation is used for training, which reduces the dependence on large-scale external labeled data and high-quality labels, thereby enhancing the network's generalization ability and robustness.
It improves the training efficiency and recognition performance of remote sensing image processing models, reduces computational costs, and is suitable for resource-constrained practical application scenarios.
Smart Images
Figure CN119314063A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of SAR image recognition, and in particular to a SAR image recognition method based on self-knowledge distillation. Background Art
[0002] SAR (Synthetic Aperture Radar) has the advantages of penetrating clouds and fog and imaging at night, showing its superiority over optical remote sensors. It has been rapidly developed in the 1990s, especially the development of polarization radar and interferometric radar technology, and has been widely used in aviation, aerospace, weapons, electronics, ships, mapping, defense, emergency response, environmental protection and other industries and fields. SAR technology has unique advantages in disaster monitoring, environmental monitoring, ocean monitoring, resource exploration, crop yield estimation, mapping and military applications, and can play a role that other remote sensing methods cannot play, and is increasingly valued by countries around the world.
[0003] In the application of SAR technology, image processing is one of the key links. At present, SAR image recognition usually relies on traditional machine learning methods or primary deep learning models, which often require a lot of computing resources and complex model adjustments. When facing edge devices, due to the complexity of the network, when processing SAR images, it often requires more space and performance to deploy. Summary of the invention
[0004] Based on this, it is necessary to provide a SAR image recognition method based on self-knowledge distillation to address the above technical problems.
[0005] A SAR image recognition method based on self-knowledge distillation includes the following steps:
[0006] Acquire an optical remote sensing image and a SAR image, and generate an optical transformation image according to the optical remote sensing image and the SAR image;
[0007] Build a self-knowledge distillation recognition model and initialize network model parameters;
[0008] sequentially inputting the optical remote sensing image, the optical transformation image and the SAR image into the self-knowledge distillation recognition model, training the self-knowledge distillation recognition model, and obtaining a target recognition network;
[0009] The SAR image to be identified is obtained and input into the target recognition network to obtain the target recognition result.
[0010] In one embodiment, generating an optically transformed image according to the optical remote sensing image and the SAR image comprises:
[0011] The optical transformation image is generated by the following formula:
[0012]
[0013] Among them, d ij represents the optical transformation image, r ij Represented as the optical transformation image label vector, g i represents the optical remote sensing image, s j represents the SAR image, e and f represent the weighting factors, m i Represents the true label vector corresponding to the optical remote sensing image, n j Represents the true label vector corresponding to the SAR image.
[0014] In one embodiment, the self-knowledge distillation recognition model includes:
[0015] Feature extractors and classifiers.
[0016] In one embodiment, the optical remote sensing image, the optical transformation image and the SAR image are sequentially input into the self-knowledge distillation recognition model, and the self-knowledge distillation recognition model is trained to obtain a target recognition network, including:
[0017] Input the input image into the self-knowledge distillation recognition model, calculate the cross entropy loss and the knowledge distillation loss in combination with a preset loss function; calculate the gradient of the network parameters according to the cross entropy loss and the knowledge distillation loss, and update the network model parameters to obtain the target recognition network;
[0018] Wherein, the optical remote sensing image, the optical transformation image and the SAR image are inputted as input images in sequence.
[0019] In one embodiment, inputting the input image into the self-knowledge distillation recognition model, combining a preset loss function, and calculating the cross entropy loss and the knowledge distillation loss includes:
[0020] Inputting the input image into the feature extractor to obtain a feature extraction vector;
[0021] According to the feature extraction vector and the classifier, combined with a preset loss function, a cross entropy loss is obtained;
[0022] According to the feature extraction vector and the classifier, combined with a preset loss function, a knowledge distillation loss is obtained.
[0023] In one embodiment, the preset loss function includes:
[0024] Cross entropy loss function, Dropout function and regularization loss function.
[0025] In one embodiment, according to the feature extraction vector and the classifier, combined with a preset loss function, the cross entropy loss is obtained, which includes:
[0026] The feature extraction vector is input into the classifier to obtain the logit value:
[0027]
[0028] Among them, z represents the logit value, represents the neural network model determined by the parameter θ, f(x) represents the feature extraction vector, and h(·) represents the classifier;
[0029] Get the classification category corresponding to the real SAR image, and get the logit category according to the logit value:
[0030]
[0031] Among them, p represents the logit category probability; T represents the distillation temperature coefficient, n represents the logit category, z l represents the lth category of logit, z j represents the jth category of logit, M represents the total category of logit; exp(·) represents the network output logical unit; x represents the input image; represents the input image group; y represents the classification category corresponding to the input image. represents the classification category group that the input image actually corresponds to, and N represents the total classification category that the input image actually corresponds to;
[0032] The cross entropy loss function is used to calculate the cross entropy loss between the logit category probability and the classification category actually corresponding to the input image.
[0033] In one embodiment, according to the feature extraction vector and the classifier, combined with a preset loss function, the knowledge distillation loss is obtained, which includes:
[0034] Apply the Dropout function on the feature extraction vector to randomly close some features and obtain two sampled sampled feature vectors:
[0035]
[0036] a j ~Bernoulli(β)
[0037] in, Represents the sampled feature vector, including: and a represents the eigenvector, ⊙ represents the element product, a jrepresents the j-th feature vector, β represents the inactivation rate, f(x) represents the feature extraction vector, h(·) represents the classifier, and Bernoulli(·) represents the Bernoulli distribution;
[0038] For each sampled feature vector, pass it through the classifier in turn to obtain two sets of posterior probabilities p(y|x; u,θ,T) and p(y|x; v,θ,T);
[0039] The knowledge distillation loss is calculated based on two sets of posterior probabilities and the regularization loss function:
[0040]
[0041] in, represents the knowledge distillation loss, represents the relative entropy algorithm, p(n|x;u,θ,T) represents the sampling feature vector The corresponding posterior probability, p(n|x; v,θ,T) represents the sampled feature vector The corresponding posterior probability.
[0042] In one embodiment, the gradient of the network parameters is calculated according to the cross entropy loss and the knowledge distillation loss, and the network model parameters are updated to obtain the target recognition network, including:
[0043] The total loss is calculated based on the cross entropy loss and the knowledge distillation loss:
[0044]
[0045] in, represents the total loss, represents the cross entropy loss, λ represents the weight hyperparameter, T represents the distillation temperature coefficient, represents the knowledge distillation loss;
[0046] The gradient of the network parameters is calculated according to the total loss, and the network model parameters are updated to obtain the target recognition network.
[0047] Compared with the prior art, the advantages and beneficial effects of the present invention are: the present invention significantly improves the training efficiency and recognition performance of the remote sensing image processing model through the strategy of endogenous self-knowledge distillation. Compared with the traditional teacher-student framework, the present invention does not rely on pre-trained large models, but uses random inactivation technology to generate multiple sub-models within a single network during the training process, and transfers knowledge between these sub-models. Such a self-knowledge distillation process reduces the dependence on large-scale external annotated data and reduces the demand for high-quality labels, thereby solving the problem of remote sensing image recognition tasks in data-scarce environments. The method of the present invention effectively utilizes computing resources and avoids the large amount of computing time and cost required to train multiple separate networks. In addition, the knowledge exchange and fusion of the internal paths of the model enhance the generalization ability and robustness of the network for synthetic aperture radar (SAR) images. Due to its flexibility and adaptability, the present invention can be easily applied to different network architectures and SAR image recognition tasks of various scales. The combined effect of these advantages not only improves recognition efficiency and accuracy, but also reduces implementation costs, enabling deep learning technology to be widely used in practical application scenarios with limited resources. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 The figure is a flow chart of a SAR image recognition method based on self-knowledge distillation in one embodiment. DETAILED DESCRIPTION
[0049] Before describing the specific embodiments of the present invention, the overall concept of the present invention is described as follows:
[0050] The present invention is mainly developed for the SAR image recognition process. At present, there are problems in the field of SAR image classification, such as high-performance models are often large in size, the traditional single-teacher model distillation information transmission efficiency is low, and the differences in model structure and parameters make it difficult to select an excellent teacher model. When the amount of remote sensing data is limited, the distillation training effect is not satisfactory.
[0051] Therefore, the present invention proposes a SAR image recognition method based on self-knowledge distillation. Self-knowledge distillation utilizes a neural network model through a discarding algorithm. This model usually contains a series of feature extraction layers and a fully connected layer for classification tasks. In the model setting stage, this structure ensures that the algorithm can be implemented on different network architectures, thereby increasing its flexibility and applicability.
[0052] After introducing the overall concept of the present invention, in order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below through specific implementation methods in conjunction with the accompanying drawings.
[0053] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in one or more embodiments of this specification should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in one or more embodiments of this specification do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.
[0054] For ease of understanding, the terms involved in the embodiments of the present invention are explained below:
[0055] SAR image: Synthetic Aperture Radar (SAR) image, is a type of image used for remote sensing mapping. SAR is an active microwave sensor that generates images by emitting pulsed electromagnetic waves to the ground and receiving reflected signals to obtain high-resolution images of the target. SAR images can not only view the terrain features of the ground and monitor land use changes, but can also be used in many fields such as hydrological monitoring and resource development planning.
[0056] In one embodiment, Figure 1 As shown, a SAR image recognition method based on self-knowledge distillation is provided, comprising the following steps:
[0057] Step S101: acquiring an optical remote sensing image and a SAR image, and generating an optical transformation image according to the optical remote sensing image and the SAR image.
[0058] Specifically, an optical remote sensing image and a SAR image data set are obtained, wherein the SAR image data set includes multiple SAR images, and an optical transformation image is generated according to the optical remote sensing image and the SAR image.
[0059] On this basis, generating an optical transformation image according to the optical remote sensing image and the SAR image includes:
[0060] The optical transformation image is generated by the following formula:
[0061]
[0062] Among them, d ij represents the optical transformation image, rij Represented as the optical transformation image label vector, g i represents the optical remote sensing image, s j represents the SAR image, e and f represent the weighting factors, m i Represents the true label vector corresponding to the optical remote sensing image, n j Represents the true label vector corresponding to the SAR image.
[0063] Specifically, an optical transformation image is generated through optical remote sensing images and SAR images, which can reduce the problem of insufficient model performance caused by lack of SAR image data and the problem of differences between optical and SAR images, and enhance the performance of the SAR recognition model.
[0064] Step S102, constructing a self-knowledge distillation recognition model and initializing network model parameters.
[0065] On this basis, the self-knowledge distillation recognition model includes:
[0066] Feature extractors and classifiers.
[0067] Specifically, the self-knowledge distillation recognition model uses a neural network model through a discard algorithm, which usually contains a series of feature extractors (feature extraction convolutional layers) and a classifier (fully connected layers) for classification tasks. In the model setting stage, this structure ensures that the algorithm can be implemented on different network architectures, thereby increasing its flexibility and applicability.
[0068] Initialize the network model parameters, define the global feature vector of the SAR image x as f(x), and the final fully connected layer of the network as h(·). Set the hyperparameters, especially the self-knowledge distillation weight λ.
[0069] Step S103, sequentially inputting the optical remote sensing image, the SAR image and the optical transformation image into the self-knowledge distillation recognition model, training the self-knowledge distillation recognition model, and obtaining a target recognition network.
[0070] Specifically, optical remote sensing images, optical transformation images and SAR images are acquired and sequentially input into the self-knowledge distillation recognition model for model training.
[0071] On this basis, the optical remote sensing image, the optical transformation image and the SAR image are sequentially input into the self-knowledge distillation recognition model, and the self-knowledge distillation recognition model is trained to obtain a target recognition network including:
[0072] Input the input image into the self-knowledge distillation recognition model, calculate the cross entropy loss and the knowledge distillation loss in combination with a preset loss function; calculate the gradient of the network parameters according to the cross entropy loss and the knowledge distillation loss, and update the network model parameters to obtain the target recognition network;
[0073] Wherein, the optical remote sensing image, the optical transformation image and the SAR image are inputted as input images in sequence.
[0074] Specifically, the optical remote sensing image, the optical transformation image, and the SAR image are sequentially input into the self-knowledge distillation recognition model, and the cross entropy loss and the knowledge distillation loss are calculated in combination with a preset loss function to perform model training. On this basis, the input image is input into the self-knowledge distillation recognition model, and the cross entropy loss and the knowledge distillation loss are calculated in combination with a preset loss function, including:
[0075] Inputting the input into the feature extractor to obtain a feature extraction vector;
[0076] According to the feature extraction vector and the classifier, combined with a preset loss function, a cross entropy loss is obtained;
[0077] According to the feature extraction vector and the classifier, combined with a preset loss function, a knowledge distillation loss is obtained.
[0078] Specifically, get the data batch input group Each image in the input image set is x, For a given input x, we first calculate its global feature vector f(x) after passing through the feature extractor.
[0079] Then, the global feature vector is passed through the classifier and combined with the preset loss function to obtain the cross entropy loss.
[0080] A dropout layer is integrated after all feature extraction layers. The main function of this layer is to randomly disable certain neurons in the network during training. This random dropout operation helps to simulate multiple different network configurations, each with slightly different weights. By implementing the dropout operation, two different sampled feature vectors u⊙f(x) and v⊙f(x) are sampled from f(x), where u and v are independently sampled from a Bernoulli distribution with a specified dropout rate β. Two sets of sampled feature vectors representing the same input but observed from different perspectives are generated. In this way, the network is able to share knowledge between different model variants and enhance the generalization ability of the model.
[0081] The two sampled feature vectors are then fed into the classifier to generate the corresponding output logarithmic values, which are converted into the posterior probability distribution of the class through the softmax function to obtain the knowledge distillation loss.
[0082] On this basis, the loss function includes:
[0083] Cross entropy loss function, Dropout function and regularization loss function.
[0084] Specifically, the loss function is designed by integrating the standard cross entropy loss function and the self-knowledge distillation loss function. The self-knowledge distillation loss function includes the Dropout function and the regularization loss function. The self-knowledge distillation loss function uses the hyperparameter λ to adjust its weight in the total loss. Through this design, the model not only pursues classification accuracy, but also strengthens the consistency of internal features, so that the model can show better generalization ability when facing unknown data.
[0085] On this basis, according to the feature extraction vector and the classifier, combined with the preset loss function, the cross entropy loss is obtained, including:
[0086] The feature extraction vector is input into the classifier to obtain the logit value:
[0087]
[0088] Among them, z represents the logit value, represents the neural network model determined by the parameter θ, f(x) represents the feature extraction vector, and h(·) represents the classifier;
[0089] Get the classification category corresponding to the real SAR image, and get the logit category according to the logit value:
[0090]
[0091] Among them, p represents the logit category probability; T represents the distillation temperature coefficient, n represents the logit category, z l represents the lth category of logit, z j represents the jth category of logit, M represents the total category of logit; exp(·) represents the network output logical unit; x represents the input image; represents the input image group; y represents the classification category corresponding to the input image. represents the classification category group that the input image actually corresponds to, and N represents the total classification category that the input image actually corresponds to;
[0092] The cross entropy loss function is used to calculate the cross entropy loss between the logit category probability and the classification category actually corresponding to the input image.
[0093] Specifically, each extracted feature extraction vector is passed through the classifier to obtain the predicted category logit value, and the logit of the output layer is marked as It is a neural network model determined by the parameter θ. If the logit of the category is represented as z l , then the probability of the lth category can be expressed as follows,
[0094]
[0095] The cross entropy loss function is used to calculate the cross entropy loss between the logit category probability and the classification category corresponding to the input image.
[0096] On this basis, according to the feature extraction vector and the classifier, combined with the preset loss function, the knowledge distillation loss is obtained, including:
[0097] Apply the Dropout function on the feature extraction vector to randomly close some features and obtain two sampled sampled feature vectors:
[0098]
[0099] a j ~Bernoulli(β)
[0100] in, Represents the sampled feature vector, including: and a represents the eigenvector, ⊙ represents the element product, a j represents the j-th feature vector, β represents the inactivation rate, f(x) represents the feature extraction vector, h(·) represents the classifier, and Bernoulli(·) represents the Bernoulli distribution;
[0101] For each sampled feature vector, pass it through the classifier in turn to obtain two sets of posterior probabilities p(y|x; u,θ,T) and p(y|x; v,θ,T);
[0102] The knowledge distillation loss is calculated based on two sets of posterior probabilities and the regularization loss function:
[0103]
[0104] in, represents the knowledge distillation loss, represents the relative entropy algorithm, p(y|x; u,θ,T) represents the sampling feature vector The corresponding posterior probability, p(y|x; v,θ,T) represents the sampled feature vector The corresponding posterior probability.
[0105] Specifically, SD-Dropout is applied to perform self-knowledge distillation. Dropout is applied to the feature extraction vector to randomly close some features and obtain two sampled feature vectors and This is achieved by introducing a deactivation layer after sampling the feature vector in each layer, which can be expressed as:
[0106]
[0107] in, Represents the sampled feature vector, including: and , a represents the eigenvector, ⊙ represents the element product, a j represents the j-th feature vector, β represents the inactivation rate, f(x) represents the feature extraction vector, h(·) represents the classifier, and Bernoulli(·) represents the Bernoulli distribution.
[0108] For each sampled feature vector, the classifier is used again to obtain two sets of posterior probabilities p(y|x; u,θ,T) and p(y|x; v,θ,T). The regularized loss function is used to minimize and The KL divergence between these two logit values is used to extract edge knowledge. Using the shared encoder, only the labeling operation is performed on the fully connected layer. This regularized loss function is applied in a symmetrical way. Given an input image x, representing the classification category y corresponding to the input image, and assuming that the random inactivation operation is u, v, the loss of this party can be defined as:
[0109]
[0110] During the self-knowledge distillation process, the difference between these two different posterior distributions is measured and minimized using the Kulbeck-Leibler divergence (KL-divergence). By minimizing the difference between these two distributions, the network can learn to maintain consistent class predictions under different dropout patterns, resulting in a more stable and robust learning process.
[0111] On this basis, the gradient of the network parameters is calculated according to the cross entropy loss and the knowledge distillation loss, and the network model parameters are updated to obtain the target recognition network including:
[0112] The total loss is calculated based on the cross entropy loss and the knowledge distillation loss:
[0113]
[0114] in, represents the total loss, represents the cross entropy loss, λ represents the weight hyperparameter, T represents the distillation temperature coefficient, represents the knowledge distillation loss;
[0115] The gradient of the network parameters is calculated according to the total loss, and the network model parameters are updated to obtain the target recognition network.
[0116] Specifically, the cross entropy loss and knowledge distillation loss are combined, and the total loss is calculated according to the hyperparameter λ:
[0117]
[0118] in, represents the total loss, represents the cross entropy loss, λ represents the weight hyperparameter, T represents the distillation temperature coefficient, represents the knowledge distillation loss;
[0119] Perform backpropagation, i.e. calculate the total loss The gradient of the network model parameters is calculated and updated to reduce the total loss.
[0120] Self-knowledge distillation is performed internally when training deep learning models through random dropout. The generalization ability of the model is improved by introducing dropout method changes during training to simulate multiple models and make their predictions consistent. This consistency is achieved by minimizing the KL divergence between the predicted outputs of the two dropout versions. The present invention does not require additional models or parameters to perform self-knowledge distillation.
[0121] Step S104, obtaining a SAR image to be identified, inputting it into the target recognition network, and obtaining a target recognition result.
[0122] Specifically, the SAR image to be identified is input into the trained target recognition network to obtain the target recognition result.
[0123] The present invention proposes a SAR image recognition method based on self-knowledge distillation. Based on the targeted optimization strategy and model adjustment of SAR images, this method is not only suitable for general model compression image processing tasks, but also optimized for processing SAR image classification problems in edge devices. Random deactivation technology is used to achieve self-distillation within the model. By introducing deactivation operations at multiple key levels of the model, multiple model variants are created. These variants exhibit different behaviors at different training stages, and this internal difference helps the model learn more robust feature expressions. No additional parameters or training process are required, which simplifies the traditional knowledge distillation method and optimizes the training process without external labels.
[0124] During the self-distillation process, due to the randomness of deactivation, the model presents a different perspective at each forward propagation. This mechanism enables the model to take into account different representations of the data during the learning process, thereby enhancing the depth and accuracy of the model's understanding of the input data.
[0125] The KL divergence is used to measure the consistency of the output of each variant of the model, and this measure is optimized to force different instances within the model to reach consistency and achieve self-distillation of knowledge. This design not only optimizes the output of the model, but also enhances the consistency of the internal structure of the model.
[0126] It should be noted that the method of the embodiment of the present invention can be performed by a single device, such as a computer or a server. The method of this embodiment can also be applied in a distributed scenario and completed by multiple devices cooperating with each other. In the case of such a distributed scenario, one of the multiple devices can only perform one or more steps in the method of the embodiment of the present invention, and the multiple devices will interact with each other to complete the described method.
[0127] It should be noted that some embodiments of the present invention are described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the above embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0128] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present invention (including the claims) is limited to these examples. Under the concept of the present invention, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the embodiments of the present invention as described above, which are not provided in detail for the sake of simplicity.
[0129] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code that includes one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention belong.
[0130] While the invention has been described in conjunction with specific embodiments thereof, many alternatives, modifications and variations of these embodiments will be apparent to those skilled in the art in light of the foregoing description.
[0131] The embodiments of the present invention are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present invention should be included in the protection scope of the present invention.
Claims
1. A SAR image recognition method based on self-knowledge distillation, characterized in that: include: Acquire an optical remote sensing image and a SAR image, and generate an optical transformation image according to the optical remote sensing image and the SAR image; Build a self-knowledge distillation recognition model and initialize network model parameters; sequentially inputting the optical remote sensing image, the optical transformation image and the SAR image into the self-knowledge distillation recognition model, training the self-knowledge distillation recognition model, and obtaining a target recognition network; The SAR image to be identified is obtained and input into the target recognition network to obtain the target recognition result.
2. The SAR image recognition method based on self-knowledge distillation according to claim 1, characterized in that: Generating an optical transformation image according to the optical remote sensing image and the SAR image comprises: The optical transformation image is generated by the following formula: Among them, d ij represents the optical transformation image, r ij Represented as the optical transformation image label vector, g i represents the optical remote sensing image, s j represents the SAR image, e and f represent the weighting factors, m i Represents the true label vector corresponding to the optical remote sensing image, n j Represents the true label vector corresponding to the SAR image.
3. The SAR image recognition method based on self-knowledge distillation according to claim 1, characterized in that: The self-knowledge distillation recognition model includes: Feature extractors and classifiers.
4. The SAR image recognition method based on self-knowledge distillation according to claim 3, characterized in that: The step of sequentially inputting the optical remote sensing image, the optical transformation image, and the SAR image into the self-knowledge distillation recognition model, and training the self-knowledge distillation recognition model to obtain a target recognition network comprises: Input the input image into the self-knowledge distillation recognition model, calculate the cross entropy loss and the knowledge distillation loss in combination with a preset loss function; calculate the gradient of the network parameters according to the cross entropy loss and the knowledge distillation loss, and update the network model parameters to obtain the target recognition network; Wherein, the optical remote sensing image, the optical transformation image and the SAR image are inputted as input images in sequence.
5. The SAR image recognition method based on self-knowledge distillation according to claim 4, characterized in that: Inputting the input image into the self-knowledge distillation recognition model, combining a preset loss function, and calculating the cross entropy loss and the knowledge distillation loss includes: Inputting the input image into the feature extractor to obtain a feature extraction vector; According to the feature extraction vector and the classifier, combined with a preset loss function, a cross entropy loss is obtained; According to the feature extraction vector and the classifier, combined with a preset loss function, a knowledge distillation loss is obtained.
6. The SAR image recognition method based on self-knowledge distillation according to claim 5, characterized in that: The preset loss function includes: Cross entropy loss function, Dropout function and regularization loss function.
7. The SAR image recognition method based on self-knowledge distillation according to claim 6, characterized in that: The step of obtaining the cross entropy loss based on the feature extraction vector and the classifier in combination with a preset loss function includes: The feature extraction vector is input into the classifier to obtain the logit value: Among them, z represents the logit value, represents the neural network model determined by the parameter θ, f(x) represents the feature extraction vector, and h(·) represents the classifier; Get the classification category that actually corresponds to the input image, and get the logit category based on the logit value: Among them, p represents the logit category probability; T represents the distillation temperature coefficient, n represents the logit category, z l represents the lth category of logit, z j represents the jth category of logit, M represents the total category of logit; exp(·) represents the network output logical unit; x represents the input image; represents the input image group; y represents the classification category corresponding to the input image. represents the classification category group that the input image actually corresponds to, and N represents the total classification category that the input image actually corresponds to; The cross entropy loss function is used to calculate the cross entropy loss between the logit category probability and the classification category actually corresponding to the input image.
8. The SAR image recognition method based on self-knowledge distillation according to claim 6, characterized in that: The knowledge distillation loss is obtained according to the feature extraction vector and the classifier in combination with a preset loss function, including: Apply the Dropout function on the feature extraction vector to randomly close some features and obtain two sampled sampled feature vectors: a j ~Bernoulli(b) in, Represents the sampled feature vector, including: and a represents the eigenvector, ⊙ represents the element product, a j represents the j-th feature vector, β represents the inactivation rate, f(x) represents the feature extraction vector, h(·) represents the classifier, and Bernoulli(·) represents the Bernoulli distribution; For each sampled feature vector, pass it through the classifier in turn to obtain two sets of posterior probabilities p(y|x; u,θ,T) and p(y|x; v,θ,T); The knowledge distillation loss is calculated based on two sets of posterior probabilities and the regularization loss function: in, represents the knowledge distillation loss, represents the relative entropy algorithm, p(y|x; u,θ,T) represents the sampling feature vector The corresponding posterior probability, p(y|x; v,θ,T) represents the sampled feature vector The corresponding posterior probability.
9. The SAR image recognition method based on self-knowledge distillation according to claim 1, characterized in that: The step of calculating the gradient of the network parameters according to the cross entropy loss and the knowledge distillation loss, and updating the network model parameters to obtain the target recognition network includes: The total loss is calculated based on the cross entropy loss and the knowledge distillation loss: in, represents the total loss, represents the cross entropy loss, λ represents the weight hyperparameter, T represents the distillation temperature coefficient, represents the knowledge distillation loss; The gradient of the network parameters is calculated according to the total loss, and the network model parameters are updated to obtain the target recognition network.